The zen claims that arithmetic settles

zenlm/papers had a sweeping withdrawal pass, 5deeeaa, and the owner reverted it in
f34f609 for asserting more than it verified. That objection sets the rule here: a
claim changes only where the proof is in the paper itself or in arithmetic, never
because a run could not be found. Twenty-seven of the eighty-one zen papers meet
that bar. Every correction quotes a number the paper already carries or one derived
from its own numbers by a division a reader can repeat.

Twelve papers led with a figure that appears in no row of the table it summarises.
zen-agent averaged GAIA 83.1/68.4/42.3 to 71.2 while its four baseline rows are each
the plain mean of their own three levels; the mean is 64.6. zen-embeddings-retrieval
and zen3-embedding claim MTEB averages of 72.4 and 74.3 over tables that supply their
task counts, where the 56-task weighted means are 67.1 and 67.4.
zen-mathematical-reasoning opened with five figures, none of them any row, and
explained one as best-of-32 sampling when its own sampling table puts best-of-32 at a
different number. zen-dso-protocol claimed 94% of centralized quality and a 78%
bandwidth cut against tables reading 99.4% and 99.8%-and-better, the second of them
the table that claim captions. zen-synthetic-data claimed an 8.4-point MT-Bench gain
on a scale that runs to 10, from a 7.84 baseline, where its table says 0.88.
zen-3d quoted the Chair column as the ShapeNet result. zen-multilingual understated
its own XNLI average. Also zen-hallucination-reduction, zen-distributed-training,
zen-finetuning, zen-legal-ai and zen-coder-flash.

Eight more are settled by arithmetic against the hardware or the parameter count.
zen-nano, zen-eco and zen-scribe each report INT4 needing as much memory as FP16 or
more, while their own FP16 and INT8 rows are exactly two and one bytes per parameter;
zen_family_overview already carried the right values. The two zen-designer papers size
a stated 235B model at 220 GB in FP16, which is a 110B model, and zen-quantization puts
the same model at 470 GB. zen-omni and zen-next recommend cards smaller than the memory
the same row requires. zen-live captions a latency table to an A10G when its own
deployment table assigns that 98 ms to the H100. zen-hardware-optimization calls a
throughput table FP8+INT4 when its cumulative table reaches the identical 31,200 tok/s
through INT4 + FP16.

Six papers share one template whose visual benchmark block bundles two metrics
computed over images a model generates with one computed over text a model emits
about an image. No model produces all three, and the block went to four papers across
four modalities, so each received at least one metric it cannot compute: VQA v2 and a
token rate for two diffusion models, FID and CLIP score for two vision-language models.
Each now keeps the rows its own stated output type can produce.

Throughput is the one category where the defect is the missing condition rather than
the number. Decode reads the active weight set once per token, so a single stream
cannot outrun bandwidth over footprint, and zen4-mini, zen3-nano and zen-coder-flash
all report above their own bound. Batched serving would make those figures legal and
none of the three records a batch size, so the numbers stand and each table now carries
its bound and the arithmetic behind it. AIME gets the same treatment in
zen-mathematical-reasoning: a score that is not a multiple of 1/30 on a 30-problem exam
is legal under averaging, and the caption now says which row has its sample count on
record and which do not.

What was examined and left is in LLM.md with the reason in each case. Training fleets
absent from this estate prove nothing about models trained on rented capacity.
zen-reranker's 98% semantic preservation is uncited rather than contradicted.
zen-safety-evaluation and zen-vision-architecture state no weighting, so their averages
are not determined. zen4-coder-flash puts FP8 on an A100, and Ampere has no FP8 tensor
cores, but its claims there are footprint claims and FP8 weight-only storage does run
on Ampere.

All twenty-seven rebuild under latexmk -Werror. Nineteen have diverged from their
counterparts in zenlm/papers and six exist only here, so those fixes do not reach the
other site. zen-3d and zen-agent were byte-identical across the two repositories until
now, and are the two to mirror first.

Co-authored-by: Hanzo Dev <[email protected]>

LLM.md

@@ -136,16 +136,16 @@ short section saying what was removed and why. Two are different: `pubsub` and
 so those tables stay with their provenance stated. The gap there is a run, not a
 mechanism.
 
-**The eighty-one `zen/` papers are not corrected and need a decision.** They hold
-321 of the 353 defects, in three recurring shapes: training and inference
-attributed to H100 and A100 fleets of 8 to 16,384 GPUs; efficiency, energy and
-carbon figures with no baseline named; and abstracts whose headline number
-contradicts the paper's own table. Some are checkable without any estate fact —
-decode throughput above the memory-bandwidth ceiling of the part it names, FP8 on
-Ampere, AIME scores that are not multiples of 1/30 on a 30-problem exam, INT4
-requiring more memory than FP16.
+**The eighty-one `zen/` papers hold the other 321,** in three recurring shapes:
+training and inference attributed to H100 and A100 fleets of 8 to 16,384 GPUs;
+efficiency, energy and carbon figures with no baseline named; and abstracts whose
+headline number contradicts the paper's own table. Some are checkable without any
+estate fact — decode throughput above the memory-bandwidth ceiling of the part it
+names, FP8 on Ampere, AIME scores that are not multiples of 1/30 on a 30-problem
+exam, INT4 requiring more memory than FP16. That checkable subset is now corrected;
+see below.
 
-They are left alone deliberately. `zenlm/papers` already had a sweeping
+The rest are left alone deliberately. `zenlm/papers` already had a sweeping
 withdrawal pass, `5deeeaa`, which the owner reverted in `f34f609`: *"That commit
 asserted it; nothing verified it... a paper that disclaims its own results is
 worse than one that under-cites them. Where these papers are thin it is in
@@ -158,6 +158,82 @@ what the rest could be: `zen-mixture-of-experts.tex` states plainly that it
 publishes no number it has not measured and names its harness, and
 `zen-coder_whitepaper.tex` reports no scores and labels its costs as estimates.
 
+## The provable subset of the zen papers
+
+A second pass corrected twenty-seven of the eighty-one, on one rule: change a claim
+only where the proof is in the paper itself or in arithmetic, never because a run
+could not be found. Every correction quotes a number the paper already carries, or
+one derived from the paper's own numbers by a division a reader can repeat.
+
+**A figure that appears in no row of the table it summarises.** Twelve papers.
+`zen-agent` averaged GAIA 83.1/68.4/42.3 to 71.2 while its four baseline rows are
+the plain mean of their own three levels; the mean is 64.6. `zen-embeddings-retrieval`
+and `zen3-embedding` state MTEB averages of 72.4 and 74.3 over tables that supply
+their task counts, where the 56-task weighted means are 67.1 and 67.4.
+`zen-mathematical-reasoning` opened with five figures, none of which was any row,
+and explained one of them as best-of-32 when its own sampling table puts best-of-32
+elsewhere. `zen-dso-protocol` claimed 94% of centralized quality and a 78% bandwidth
+cut against tables reading 99.4% and 99.8%-and-better. Also `zen-hallucination-reduction`
+(reductions attached to the wrong benchmarks, one attached to nothing),
+`zen-synthetic-data` (an 8.4-point MT-Bench gain on a 1-to-10 scale from a 7.84
+baseline, where the table says 0.88), `zen-distributed-training`, `zen-finetuning`,
+`zen-legal-ai`, `zen-coder-flash`, `zen-3d` (the abstract quoted the Chair column as
+the ShapeNet result) and `zen-multilingual`, whose XNLI average was understated.
+
+**Arithmetic that the silicon or the scale forbids.** `zen-nano`, `zen-eco` and
+`zen-scribe` each report INT4 needing as much memory as FP16 or more, while their own
+FP16 and INT8 rows are exactly two and one bytes per parameter; `zen_family_overview`
+already carried the right values. Both `zen-designer` papers size a stated 235B model
+at 220 GB in FP16, which is a 110B model; `zen-quantization` puts the same model at
+470 GB. `zen-omni` and `zen-next` recommend cards smaller than the memory the same
+row requires. `zen-live` captions a latency table to an A10G when its own deployment
+table assigns that 98 ms to the H100 and gives the A10G 113 ms.
+`zen-hardware-optimization` calls a throughput table FP8+INT4 when its own cumulative
+table reaches the identical 31,200 tok/s figure through INT4 + FP16.
+
+**A metric that cannot describe the paper's subject.** Six papers share one template
+whose "Visual Understanding Benchmarks" block bundles two metrics computed over images
+a model generates (FID, CLIP score) with one computed over text a model emits about an
+image (VQA v2). No model produces all three. The block went to four papers across four
+modalities, so each got at least one metric it cannot compute: the diffusion papers
+`zen-artist` and `zen-artist-edit` reported VQA v2 and a token rate, and the
+vision-language papers `zen-designer-instruct` and `zen-designer-thinking` reported FID
+and CLIP score. Each paper now keeps the rows its own stated output type can produce.
+
+**Throughput, where the defect is the missing condition rather than the number.**
+Decode reads the active weight set once per token, so one stream cannot outrun
+bandwidth over footprint. `zen4-mini`, `zen3-nano` and `zen-coder-flash` report figures
+above their own bound, but batched serving would make them legal and none of the three
+records a batch size. The numbers stand; each table now carries its bound and the
+arithmetic behind it. This is the same treatment AIME gets in
+`zen-mathematical-reasoning`, where a score that is not a multiple of 1/30 is legal
+under averaging and the caption now says so.
+
+**Left alone, and why.** Training fleets absent from this estate prove nothing about
+models trained on rented capacity. `zen-reranker`'s 98% semantic preservation appears
+in no table, but the only "preserved" figure in the paper describes a different model
+at a different dimension, so it is uncited rather than contradicted.
+`zen-safety-evaluation`'s attack-vector average does not match the mean of its rows,
+but the table states no per-vector counts, so no weighting is determined.
+`zen-vision-architecture`'s MMMU overall covers 30 topics while listing 19, so the
+average is not determined either. `zen4-coder-flash` puts FP8 on an A100: Ampere has
+no FP8 tensor cores, but its claims there are footprint claims, and FP8 weight-only
+storage with dequantisation does run on Ampere.
+
+Two arithmetic problems are real and were not corrected, because fixing them means
+choosing which of several numbers is wrong. The template's pretraining sentence claims
+7 trillion tokens over 60 days on 128 A100 for a 22B-active model, which needs about
+4.5 times the peak BF16 throughput of the hardware it names before any utilisation
+discount; the same sentence in `zen-scribe` implies about 6% utilisation. And the
+derived tables in the two embedding papers quote the corrected MTEB average as their
+uncompressed baseline, so their compression and Matryoshka columns still sit above it
+and need the authors' per-configuration data to re-base.
+
+Of the twenty-seven corrected here, nineteen have diverged from `zenlm/papers`,
+six exist only in this repository, and two — `zen-3d.tex` and `zen-agent.tex` — were
+byte-identical to their counterparts there before this pass and now are not. Those two
+are the ones worth mirroring first.
+
 Note also that 69 of the 86 `zen/*.tex` here have diverged from their
 counterparts in `zenlm/papers`, so the two sites publish different text under the
 same titles. A fix applied in one place does not reach the other.

zen/zen-3d.tex

@@ -53,7 +53,7 @@
 We present \textbf{Zen-3D}, an 8-billion parameter model specialized for 3D scene understanding,
 mesh generation, and point cloud processing. Built on the Zen MoDE (Mixture of Distilled Experts)
 architecture with domain-adapted 3D tokenization, Zen-3D achieves state-of-the-art results across
-all major 3D benchmarks: 95.2\% IoU on ShapeNet, 88.3\% mIoU on ScanNet semantic segmentation,
+all major 3D benchmarks: 92.2\% mean IoU on ShapeNet, 88.3\% mIoU on ScanNet semantic segmentation,
 and 94.1\% classification accuracy on ModelNet40. Unlike prior work that treats 3D understanding
 as a downstream task grafted onto language models, Zen-3D natively encodes spatial geometry through
 a novel triplane-based 3D tokenizer and dedicated attention mechanisms for volumetric context.
@@ -439,7 +439,7 @@ tokenization, eliminating the projection bottleneck.
 
 Zen-3D demonstrates that a single 8B parameter model, trained with the Zen MoDE architecture
 and a purpose-built triplane tokenizer, can simultaneously achieve state-of-the-art results
-across 3D segmentation (ScanNet mIoU 88.3\%), reconstruction (ShapeNet IoU 95.2\%), and
+across 3D segmentation (ScanNet mIoU 88.3\%), reconstruction (ShapeNet mean IoU 92.2\%), and
 classification (ModelNet40 94.1\%). The unified backbone enables cross-task transfer that
 benefits all three capabilities simultaneously.
 

zen/zen-agent.tex

@@ -54,7 +54,7 @@ We present \textbf{Zen-Agent}, a 32-billion parameter model trained for autonomo
 multi-step planning, and compositional tool use. Zen-Agent is built on the Zen MoDE (Mixture of
 Distilled Experts) architecture and fine-tuned on the Zen Agentic Trajectory Dataset---a corpus
 of 12.4 million complete task trajectories spanning software engineering, web navigation, data
-analysis, and scientific reasoning. The model achieves 71.2\% on GAIA (Level 1--3 average),
+analysis, and scientific reasoning. The model achieves 64.6\% on GAIA (Level 1--3 average),
 42.8\% on SWE-bench Verified, and 88.3\% on HumanEval, establishing a new frontier for open
 agentic models at this parameter scale. Zen-Agent introduces \textbf{Hierarchical Task
 Decomposition (HTD)}: an explicit planning module that generates subgoal trees before executing
@@ -273,7 +273,7 @@ GPT-4o (2024-11) & 73.2 & 52.1 & 28.4 & 51.2 \\
 Claude 3.5 Sonnet & 77.3 & 56.2 & 31.7 & 55.1 \\
 Gemini 1.5 Pro & 70.1 & 48.3 & 26.1 & 48.2 \\
 Open-source SoTA (32B) & 52.7 & 31.4 & 14.2 & 32.8 \\
-\textbf{Zen-Agent 32B} & \textbf{83.1} & \textbf{68.4} & \textbf{42.3} & \textbf{71.2} \\
+\textbf{Zen-Agent 32B} & \textbf{83.1} & \textbf{68.4} & \textbf{42.3} & \textbf{64.6} \\
 \bottomrule
 \end{tabular}
 \caption{GAIA Benchmark Results (\%, higher is better)}
@@ -425,7 +425,7 @@ training at 32B scale with the Zen MoDE architecture.
 Zen-Agent demonstrates that agentic capability can be substantially improved by training on
 real task trajectories at scale, enforcing explicit hierarchical planning before action
 execution, and natively integrating with tool ecosystems via MCP. The resulting model achieves
-71.2\% GAIA and 42.8\% SWE-bench Verified, establishing a new state of the art for open
+64.6\% GAIA and 42.8\% SWE-bench Verified, establishing a new state of the art for open
 agentic models at 32B scale while using 30\% fewer tool calls than comparable methods.
 
 Future work will extend Zen-Agent to multi-agent collaboration (where multiple Zen-Agent

zen/zen-artist-edit_whitepaper.tex

@@ -128,7 +128,6 @@ Specialized attention mechanisms optimized for image editing \&   inpainting.
 \toprule
 \textbf{Benchmark} \ \ &   \textbf{Score} \\
 \midrule
-VQA v2 \ \ &   91.2\% \\
 DesignBench \ \ &   87.3\% \\
 CLIP Score \ \ &   86.6\% \\
 FID Score \ \ &   72.6 \\
@@ -137,6 +136,9 @@ FID Score \ \ &   72.6 \\
 \caption{Visual Understanding Benchmarks}
 \end{table}
 
+VQA v2 is not reported. It scores natural-language answers about an input image,
+and this model takes text and returns pixels.
+
 \subsection{Efficiency Metrics}
 
 \begin{table}[H]
@@ -145,15 +147,18 @@ FID Score \ \ &   72.6 \\
 \toprule
 \textbf{Metric} \ \ &   \textbf{Value} \\
 \midrule
-Inference Speed \ \ &   180 tokens/sec \\
 Memory Usage (INT4) \ \ &   3.5 GB \\
 Energy Efficiency \ \ &   93\% reduction \\
-Latency (First Token) \ \ &   45 ms \\
 \bottomrule
 \end{tabular}
 \caption{Efficiency Metrics}
 \end{table}
 
+A diffusion sampler emits images rather than tokens, so neither a token rate nor a
+time-to-first-token applies here; the corresponding measure is seconds per image at
+the sampler's step count.
+
+
 \section{Training Methodology}
 
 \subsection{Dataset}

zen/zen-artist_whitepaper.tex

@@ -130,7 +130,6 @@ Specialized attention mechanisms optimized for text-to-image generation.
 \toprule
 \textbf{Benchmark} & \textbf{Score} \\
 \midrule
-VQA v2 & 88.5\% \\
 DesignBench & 82.4\% \\
 CLIP Score & 84.1\% \\
 FID Score & 73.5 \\
@@ -139,6 +138,9 @@ FID Score & 73.5 \\
 \caption{Visual Understanding Benchmarks}
 \end{table}
 
+VQA v2 is not reported. It scores natural-language answers about an input image,
+and this model takes text and returns pixels.
+
 \subsection{Efficiency Metrics}
 
 \begin{table}[H]
@@ -147,15 +149,18 @@ FID Score & 73.5 \\
 \toprule
 \textbf{Metric} & \textbf{Value} \\
 \midrule
-Inference Speed & 160 tokens/sec \\
 Memory Usage (INT4) & 4 GB \\
 Energy Efficiency & 93\% reduction \\
-Latency (First Token) & 50 ms \\
 \bottomrule
 \end{tabular}
 \caption{Efficiency Metrics}
 \end{table}
 
+A diffusion sampler emits images rather than tokens, so neither a token rate nor a
+time-to-first-token applies here; the corresponding measure is seconds per image at
+the sampler's step count.
+
+
 \section{Training Methodology}
 
 \subsection{Dataset}

zen/zen-coder-flash_whitepaper.tex

@@ -36,7 +36,7 @@ We present \textbf{Zen-Coder-Flash}, an 8 billion parameter code completion mode
 from Zen-Code (14B) using the Zen MoDE (Mixture of Distilled Experts) distillation pipeline.
 Zen-Coder-Flash targets IDE-grade latency requirements: P95 time-to-first-token below 50ms
 for typical autocomplete prompts, with throughput of 850 tokens/second on a single A100 GPU.
-Despite its reduced parameter count, Zen-Coder-Flash achieves HumanEval 81.4\%---84.4\% of
+Despite its reduced parameter count, Zen-Coder-Flash achieves HumanEval 81.4\%---93.4\% of
 Zen-Code's performance at 55\% of the parameter count and 2.1$\times$ lower inference latency.
 The model supports fill-in-the-middle (FIM), 32K context, and all 80$+$ languages from
 Zen-Code, making it the recommended choice for real-time developer tooling where sub-50ms
@@ -280,7 +280,10 @@ Zen-Coder-Flash's P95 of 43ms satisfies the $<$50ms IDE latency budget.
 
 \begin{table}[H]
 \centering
-\caption{Generation throughput (tokens/second) on single A100-80GB.}
+\caption{Generation throughput (tokens/second) on single A100-80GB. The Batch=1
+column exceeds what one stream can read: BF16 weights are 16.4 GB for the 8B model and
+28.8 GB for the 14B, and the A100's 2{,}039 GB/s bounds single-stream decode at 124 and
+71 tokens per second respectively. The batched columns carry no such conflict.}
 \label{tab:throughput}
 \begin{tabular}{lccc}
 \toprule

zen/zen-designer-instruct_whitepaper.tex

@@ -139,13 +139,14 @@ Advanced reasoning through extended thinking tokens (up to 512K), enabling:
 \midrule
 VQA v2 & 95.8\% \\
 DesignBench & 92.1\% \\
-CLIP Score & 91.0\% \\
-FID Score & 71.3 \\
 \bottomrule
 \end{tabular}
 \caption{Visual Understanding Benchmarks}
 \end{table}
 
+CLIP score and FID are not reported. Both are computed over images a model
+generates, and this model takes an image and returns text.
+
 \subsection{Efficiency Metrics}
 
 \begin{table}[H]
@@ -155,7 +156,7 @@ FID Score & 71.3 \\
 \textbf{Metric} & \textbf{Value} \\
 \midrule
 Inference Speed & 25 tokens/sec \\
-Memory Usage (INT4) & 55 GB \\
+Memory Usage (INT4) & 118 GB \\
 Energy Efficiency & 90\% reduction \\
 Latency (First Token) & 180 ms \\
 \bottomrule
@@ -257,9 +258,9 @@ The model has been developed with careful attention to:
 \toprule
 \textbf{Precision} & \textbf{Memory} & \textbf{Recommended Hardware} \\
 \midrule
-FP16 & 220 GB & 4x A100 80GB \\
-INT8 & 110 GB & 2x A100 80GB \\
-INT4 & 55 GB & A100 80GB \\
+FP16 & 470 GB & 6x A100 80GB \\
+INT8 & 235 GB & 3x A100 80GB \\
+INT4 & 118 GB & 2x A100 80GB \\
 \bottomrule
 \end{tabular}
 \caption{Hardware Requirements by Precision}

zen/zen-designer-thinking_whitepaper.tex

@@ -139,13 +139,14 @@ Advanced reasoning through extended thinking tokens (up to 2M), enabling:
 \midrule
 VQA v2 & 96.3\% \\
 DesignBench & 94.2\% \\
-CLIP Score & 91.5\% \\
-FID Score & 71.1 \\
 \bottomrule
 \end{tabular}
 \caption{Visual Understanding Benchmarks}
 \end{table}
 
+CLIP score and FID are not reported. Both are computed over images a model
+generates, and this model takes an image and returns text.
+
 \subsection{Efficiency Metrics}
 
 \begin{table}[H]
@@ -155,7 +156,7 @@ FID Score & 71.1 \\
 \textbf{Metric} & \textbf{Value} \\
 \midrule
 Inference Speed & 25 tokens/sec \\
-Memory Usage (INT4) & 55 GB \\
+Memory Usage (INT4) & 118 GB \\
 Energy Efficiency & 90\% reduction \\
 Latency (First Token) & 180 ms \\
 \bottomrule
@@ -257,9 +258,9 @@ The model has been developed with careful attention to:
 \toprule
 \textbf{Precision} & \textbf{Memory} & \textbf{Recommended Hardware} \\
 \midrule
-FP16 & 220 GB & 4x A100 80GB \\
-INT8 & 110 GB & 2x A100 80GB \\
-INT4 & 55 GB & A100 80GB \\
+FP16 & 470 GB & 6x A100 80GB \\
+INT8 & 235 GB & 3x A100 80GB \\
+INT4 & 118 GB & 2x A100 80GB \\
 \bottomrule
 \end{tabular}
 \caption{Hardware Requirements by Precision}

zen/zen-distributed-training.tex

@@ -29,11 +29,11 @@ We present Decentralized Semantic Optimization (DSO), the distributed training
 infrastructure used for the Zen model family. DSO addresses the challenge of
 coordinating gradient aggregation across heterogeneous, geographically distributed
 compute nodes without a central parameter server. Key contributions include
-semantic-aware gradient compression that preserves model-semantic fidelity at 94\%
+semantic-aware gradient compression that preserves model-semantic fidelity at 98.8\%
 bandwidth reduction, a peer-to-peer gossip protocol for Byzantine-fault-tolerant
 gradient sharing, and integration with the Active Semantic Optimization (ASO) protocol
 for decentralized reward-model alignment. On a 256-node commodity cluster, DSO achieves
-convergence matching centralized training with 94\% bandwidth reduction and robustness
+convergence matching centralized training with 99.94\% bandwidth reduction and robustness
 against up to 20\% Byzantine nodes.
 \end{abstract}
 
@@ -253,9 +253,9 @@ Semantic Top-1\% + INT8 & 0.12 GB/s & 333$\times$ \\
 \caption{Bandwidth usage for 7B-parameter gradient aggregation (28B floats per step).}
 \end{table}
 
-The 94\% bandwidth reduction cited in the abstract refers to semantic vs. uncompressed
-aggregation at equivalent compression ratios. DSO default achieves 1667$\times$ vs.
-FP32 all-reduce, enabling training over 1Gb/s WAN links.
+The 98.8\% figure in the abstract is semantic Top-1\% against uncompressed FP32
+all-reduce, 0.49 GB/s against 40.0. The DSO default adds INT8 and gossip for
+1667$\times$, or 99.94\%, enabling training over 1Gb/s WAN links.
 
 \subsection{Byzantine Robustness}
 
@@ -313,7 +313,7 @@ Gossip (DSO) & $\log_f N$ & $O(f)$ & Byzantine \\
 \section{Conclusion}
 
 DSO provides a practical decentralized training infrastructure for the Zen model family,
-achieving 94\% bandwidth reduction with negligible convergence degradation ($\Delta\mathcal{L} < 0.04$)
+achieving 99.94\% bandwidth reduction with negligible convergence degradation ($\Delta\mathcal{L} \le 0.04$)
 and Byzantine fault tolerance up to 20\% malicious nodes. Integration with ASO enables
 privacy-preserving distributed reward model training. The system enables Zen training
 to run on geographically distributed, heterogeneous compute without centralized

zen/zen-dso-protocol.tex

@@ -38,8 +38,9 @@ gossip-based overlay that converges to a semantically coherent global optimum.
 We prove Byzantine fault tolerance under the assumption that at most $f < n/3$
 nodes are adversarial, derive convergence rates competitive with centralized
 training, and benchmark DSO across 4-node to 256-node deployments covering
-heterogeneous Zen MoDE model variants. DSO achieves 94\% of centralized training
-quality at 64 nodes with a 78\% reduction in inter-node communication bandwidth.
+heterogeneous Zen MoDE model variants. DSO achieves 99.4\% of centralized training
+quality at 64 fault-free nodes, and reduces inter-node communication bandwidth by
+99.8\% or better across model sizes.
 \end{abstract}
 
 \tableofcontents
@@ -70,7 +71,7 @@ Proposal ZIP-001, addresses this through:
           embedding space, allowing nodes running different Zen MoDE variants (7B, 32B,
           72B) to contribute meaningfully to a common optimization trajectory.
     \item \textbf{Communication Efficiency}: semantic gradient summaries are compressed
-          using sketching and quantization, achieving 78\% bandwidth reduction.
+          using sketching and quantization, achieving 99.8\% bandwidth reduction.
 \end{enumerate}
 
 \subsection{Threat Model}
@@ -319,8 +320,8 @@ baseline (lower = better). DSO approaches centralized quality at 64+ nodes.}
 
 \begin{table}[H]
 \centering
-\caption{Total inter-node bandwidth per training step. DSO reduces bandwidth by 78\%
-at 64 nodes vs.\ parameter-server all-reduce.}
+\caption{Total inter-node bandwidth per training step. DSO reduces bandwidth by
+99.8\% or better vs.\ parameter-server all-reduce, rising with model size.}
 \begin{tabular}{lrrrr}
 \toprule
 \textbf{Model} & \textbf{All-reduce (GB/step)} & \textbf{DSO (GB/step)} & \textbf{Reduction} \\
@@ -418,7 +419,9 @@ DSO (ZIP-001) provides a Byzantine-robust, bandwidth-efficient, and heterogeneit
 decentralized training protocol for the Zen MoDE model family. Key results:
 
 \begin{itemize}
-    \item 94--99\% of centralized training quality at 64--256 nodes.
+    \item 91.8--99.8\% of centralized training quality at 64--256 nodes,
+          the low end at the maximum tolerated Byzantine fraction.
+
     \item 99.8--99.98\% bandwidth reduction vs.\ all-reduce (model-size independent).
     \item Tolerates up to $f < n/3$ Byzantine nodes with geometric median aggregation.
     \item Semantic coherence gate provides additional defense against model poisoning.

zen/zen-eco_whitepaper.tex

@@ -154,7 +154,7 @@ HellaSwag & 71.6\% \\
 \textbf{Metric} & \textbf{Value} \\
 \midrule
 Inference Speed & 250 tokens/sec \\
-Memory Usage (INT4) & 8 GB \\
+Memory Usage (INT4) & 2 GB \\
 Energy Efficiency & 95\% reduction \\
 Latency (First Token) & 35 ms \\
 \bottomrule
@@ -258,7 +258,7 @@ The model has been developed with careful attention to:
 \midrule
 FP16 & 8 GB & RTX 3070 \\
 INT8 & 4 GB & RTX 3060 \\
-INT4 & 8 GB & M2 MacBook Air \\
+INT4 & 2 GB & M2 MacBook Air \\
 \bottomrule
 \end{tabular}
 \caption{Hardware Requirements by Precision}

zen/zen-embeddings-retrieval.tex

@@ -23,7 +23,7 @@
 \maketitle
 
 \begin{abstract}
-We present Zen Embed, a family of text embedding models derived from the Zen MoDE (Mixture of Distilled Experts) backbone, optimized for dense retrieval, semantic search, and cross-lingual understanding. Our models produce 7,680-dimensional embeddings natively compatible with BitDelta compression, reducing storage by 8$\times$ with less than 1.2\% NDCG@10 degradation. On the Massive Text Embedding Benchmark (MTEB), Zen Embed achieves an average score of 72.4 across 56 tasks. On BEIR zero-shot retrieval, we achieve 54.8 NDCG@10. On MIRACL multilingual retrieval covering 18 languages, we achieve 68.3 nDCG@10. We detail our bi-encoder training methodology, hard negative mining strategy, and cross-lingual alignment techniques.
+We present Zen Embed, a family of text embedding models derived from the Zen MoDE (Mixture of Distilled Experts) backbone, optimized for dense retrieval, semantic search, and cross-lingual understanding. Our models produce 7,680-dimensional embeddings natively compatible with BitDelta compression, reducing storage by 8$\times$ with less than 1.2\% NDCG@10 degradation. On the Massive Text Embedding Benchmark (MTEB), Zen Embed achieves an average score of 67.1 across 56 tasks. On BEIR zero-shot retrieval, we achieve 54.8 NDCG@10. On MIRACL multilingual retrieval covering 18 languages, we achieve 68.3 nDCG@10. We detail our bi-encoder training methodology, hard negative mining strategy, and cross-lingual alignment techniques.
 \end{abstract}
 
 \section{Introduction}
@@ -216,7 +216,7 @@ Retrieval & 15 & 58.3 & 55.8 & +2.5 \\
 STS & 10 & 83.4 & 81.9 & +1.5 \\
 Summarization & 1 & 31.8 & 30.4 & +1.4 \\
 \midrule
-\textbf{Average} & \textbf{56} & \textbf{72.4} & \textbf{70.2} & \textbf{+2.2} \\
+\textbf{Average} & \textbf{56} & \textbf{67.1} & \textbf{65.0} & \textbf{+2.1} \\
 \bottomrule
 \end{tabular}
 \end{table}
@@ -299,7 +299,7 @@ For billion-scale document corpora, we recommend hierarchical navigable small wo
 
 \section{Conclusion}
 
-Zen Embed delivers state-of-the-art dense retrieval quality across English and multilingual benchmarks while remaining practically deployable through BitDelta-compatible compression. The 7,680-dimensional embeddings achieve 72.4 MTEB average and 54.8 BEIR NDCG@10, and MRL compression to 2048 dimensions enables billion-scale corpora on commodity storage at 30$\times$ reduced cost.
+Zen Embed delivers state-of-the-art dense retrieval quality across English and multilingual benchmarks while remaining practically deployable through BitDelta-compatible compression. The 7,680-dimensional embeddings achieve 67.1 MTEB average and 54.8 BEIR NDCG@10, and MRL compression to 2048 dimensions enables billion-scale corpora on commodity storage at 30$\times$ reduced cost.
 
 \begin{thebibliography}{99}
 \bibitem{mteb} Muennighoff, N. et al. MTEB: Massive Text Embedding Benchmark. \textit{EACL}, 2023.

zen/zen-finetuning.tex

@@ -28,7 +28,7 @@
 We present a systematic study of fine-tuning methods for Zen models, covering
 LoRA and QLoRA parameter-efficient methods, full fine-tuning, domain adaptation,
 and catastrophic forgetting prevention. Key findings: (1) LoRA with rank 64 achieves
-96.8\% of full fine-tuning performance at 0.8\% of the trainable parameters; (2)
+98.3\% of full fine-tuning performance at 0.8\% of the trainable parameters; (2)
 optimal learning rates follow a power law in dataset size; (3) data requirements
 vary dramatically by task type (classification: 1K samples, complex generation: 50K+);
 (4) Elastic Weight Consolidation (EWC) prevents 80\% of catastrophic forgetting
@@ -327,8 +327,8 @@ of the compute cost.
 \section{Conclusion}
 
 Fine-tuning Zen models is practical and efficient across a wide range of methods
-and compute budgets. LoRA rank 64 covering attention and FFN layers achieves
-96.8\% of full fine-tuning performance at 0.8\% of parameters, making it the
+and compute budgets. LoRA rank 64 over the attention projections achieves
+98.3\% of full fine-tuning performance at 0.8\% of parameters, making it the
 default recommendation. QLoRA extends this to Zen-32B on commodity hardware.
 Full fine-tuning with EWC is warranted only when maximum task performance is
 required and a general-purpose baseline must be preserved simultaneously.

zen/zen-hallucination-reduction.tex

@@ -33,7 +33,7 @@ fine-tunes models to condition on retrieved evidence when generating factual cla
 a calibrated confidence estimate to suppress low-confidence continuations; and (3)
 \textbf{Citation-Grounded Generation (CGG)}, which trains models to emit inline
 citations linking claims to source documents. Together, these methods reduce
-hallucination rates by 61\% on TruthfulQA, 44\% on FactScore, and 53\% on HaluEval
+hallucination rates by 67\% on TruthfulQA, 53\% on FactScore, and 61\% on HaluEval
 relative to our baseline Zen MoDE-72B, with negligible perplexity degradation.
 \end{abstract}
 
@@ -418,7 +418,7 @@ hallucination at the training, decoding, and output-structure levels simultaneou
 
 We have presented a three-component system — RAFT, UAD, and CGG — that substantially
 reduces hallucination in Zen MoDE models across all tested benchmarks and domains.
-The system achieves 53--61\% hallucination reduction with less than 1\% perplexity
+The system achieves 53--67\% hallucination reduction with less than 1\% perplexity
 degradation, demonstrating that factual accuracy and fluency are not fundamentally
 in conflict. We recommend deploying all three components together for production use
 cases where factual reliability is critical.

zen/zen-hardware-optimization.tex

@@ -287,8 +287,9 @@ with only 0.2 pp MMLU degradation.
 
 \begin{table}[H]
 \centering
-\caption{End-to-end inference throughput on NVIDIA GPUs. Single GPU, FP8+INT4 mixed.
-Batch size optimized per model per GPU.}
+\caption{End-to-end inference throughput on NVIDIA GPUs. Single GPU, batch size
+optimized per model per GPU. The quantization is the INT4 + FP16 configuration of
+Table~\ref{tab:cumulative}, whose H100 MoDE-72B row is the 31,200 tok/s entry here.}
 \begin{tabular}{llrrrr}
 \toprule
 \textbf{GPU} & \textbf{Model} & \textbf{Batch} & \textbf{Tok/s} & \textbf{Latency (ms/tok)} & \textbf{Tokens/J} \\

zen/zen-legal-ai.tex

@@ -23,7 +23,7 @@
 \maketitle
 
 \begin{abstract}
-We present Zen Legal, a suite of AI models for legal research, document analysis, and contract review built on the Zen MoDE (Mixture of Distilled Experts) backbone. Zen Legal addresses the fundamental requirements of legal AI: citation-grounded generation, multi-jurisdiction reasoning, precise language preservation, and privilege protection. On standardized legal benchmarks: LegalBench 81.4\%, CUAD contract clause extraction 89.3\%, contract clause classification F1 0.923, and case outcome prediction 74.2\%. We describe the legal domain adaptation methodology, citation architecture, jurisdiction-aware reasoning, and professional responsibility guardrails.
+We present Zen Legal, a suite of AI models for legal research, document analysis, and contract review built on the Zen MoDE (Mixture of Distilled Experts) backbone. Zen Legal addresses the fundamental requirements of legal AI: citation-grounded generation, multi-jurisdiction reasoning, precise language preservation, and privilege protection. On standardized legal benchmarks: LegalBench 81.4\%, CUAD contract clause extraction F1 0.923 at 0.938 recall across 41 clause types, and case outcome prediction 74.2\%. We describe the legal domain adaptation methodology, citation architecture, jurisdiction-aware reasoning, and professional responsibility guardrails.
 \end{abstract}
 
 \section{Introduction}
@@ -280,7 +280,7 @@ Per ABA Model Rule 5.3, attorneys using AI tools must supervise the AI's work co
 
 \section{Conclusion}
 
-Zen Legal demonstrates that rigorous citation grounding, jurisdictional awareness, and professional responsibility compliance can be integrated into a production legal AI system without sacrificing capability. The 81.4\% LegalBench score, 89.3\% CUAD clause extraction, and 0.923 contract clause F1 establish Zen Legal as state-of-the-art for AI-assisted legal research, with attorney evaluations confirming practical utility comparable to junior associate work product.
+Zen Legal demonstrates that rigorous citation grounding, jurisdictional awareness, and professional responsibility compliance can be integrated into a production legal AI system without sacrificing capability. The 81.4\% LegalBench score and the 0.923 CUAD clause-extraction F1 establish Zen Legal as state-of-the-art for AI-assisted legal research, with attorney evaluations confirming practical utility comparable to junior associate work product.
 
 \begin{thebibliography}{99}
 \bibitem{legalbench} Guha, N. et al. LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models. \textit{NeurIPS}, 2023.

zen/zen-live_whitepaper.tex

@@ -186,7 +186,9 @@ Turn-taking capability is fine-tuned on a corpus of 50K annotated conversational
 
 \begin{table}[H]
 \centering
-\caption{First-Token Latency Benchmarks (single A10G GPU, FP8)}
+\caption{First-Token Latency Benchmarks (single H100 80GB GPU, FP8).
+The 98ms P95 here is the cloud configuration of Table~\ref{tab:zenlive-deploy};
+the A10G edge configuration measures 113ms.}
 \begin{tabular}{lcccc}
 \toprule
 \textbf{Model} & \textbf{P50} & \textbf{P90} & \textbf{P95} & \textbf{P99} \\
@@ -278,6 +280,7 @@ The largest gains from Zen-Live's native turn management are in timing naturalne
 \begin{table}[H]
 \centering
 \caption{Deployment Configurations}
+\label{tab:zenlive-deploy}
 \begin{tabular}{llcl}
 \toprule
 \textbf{Config} & \textbf{Hardware} & \textbf{P95 Latency} & \textbf{Concurrent Sessions} \\

zen/zen-mathematical-reasoning.tex

@@ -27,10 +27,12 @@
 \begin{abstract}
 We present a comprehensive study of mathematical reasoning capabilities in Zen models,
 covering training data construction, process reward models (PRMs) for step-level
-supervision, and a symbolic-neural hybrid verification system. Zen achieves 73.2\%
-on the MATH benchmark, 96.4\% on GSM8K, 42.1\% on AIME 2024, 78.4\% on AMC 2023,
-and 31.2\% on FrontierMath. Our step-level PRM improves MATH accuracy by 8.1 points
-over outcome-only reward models. The hybrid verification system using SymPy for
+supervision, and a symbolic-neural hybrid verification system. Zen-7B scores 71.4\%
+on the MATH benchmark, 95.8\% on GSM8K, 38.3\% on AIME 2024, 74.2\% on AMC 2023,
+and 28.1\% on FrontierMath; Zen-32B with the step-level PRM scores 83.4, 97.9, 48.2,
+84.7, and 36.9 on the same five. The PRM adds 4.8 MATH points at 7B (71.4 to 76.2)
+and 4.6 at 32B (78.8 to 83.4) over the same model without it. The hybrid
+verification system using SymPy for
 algebraic verification and neural methods for proof checking enables reliable
 answer verification across problem classes. We describe data construction, training
 protocol, and detailed performance analysis by mathematical domain.
@@ -233,11 +235,17 @@ Zen-32B & 78.8 & 97.2 & 44.3 & 81.2 & 33.4 \\
 Zen-32B + PRM & \textbf{83.4} & \textbf{97.9} & \textbf{48.2} & \textbf{84.7} & \textbf{36.9} \\
 \bottomrule
 \end{tabular}
-\caption{Mathematical reasoning benchmark results. PRM adds 4--8 points across benchmarks.}
+\caption{Mathematical reasoning benchmark results. The PRM adds between 0.7 and 4.8
+points, depending on benchmark and model size. AIME 2024 is 30 problems, so a
+single-pass score is a multiple of 3.3 points; these AIME figures are therefore
+averages over repeated samples. The sample count is on record only for
+Zen-32B + PRM ($n=128$, Section~\ref{sec:olympiad}); for the other three rows it
+is not.}
 \end{table}
 
-Note: The headline Zen-7B result of 73.2\% MATH cited in the abstract reflects the
-production model with best-of-32 sampling; single-sample accuracy is 71.4\%.
+Each row is one decode per problem on MATH and GSM8K. What test-time sampling adds
+is reported separately in Table~\ref{tab:sampling}: Zen-7B with PRM reranking reaches
+78.1\% MATH at $k=32$ and 79.0\% at $k=64$.
 
 \subsection{MATH Category Breakdown}
 
@@ -273,6 +281,7 @@ PRM (best-of-$k$) & 71.4 & 74.8 & 77.3 & 78.1 & 79.0 \\
 \bottomrule
 \end{tabular}
 \caption{MATH accuracy vs. number of samples (Zen-7B). PRM reranking outperforms majority vote.}
+\label{tab:sampling}
 \end{table}
 
 \subsection{PRM Calibration}
@@ -311,6 +320,7 @@ numerically inconsistent result) and least effective at detecting missing cases
 misinterpretation (where each individual step appears locally valid).
 
 \subsection{Olympiad-Level Performance}
+\label{sec:olympiad}
 
 On AIME 2024 (30 problems), Zen-32B with PRM-guided search and $n=128$ samples
 achieves 48.2\% (14.5/30 problems). Human AIME competitors typically score 8--15
@@ -324,9 +334,9 @@ theorem-proving capabilities beyond current neural approaches.
 
 Mathematical reasoning in Zen models benefits substantially from three components:
 domain-rich pretraining data (120B tokens), process reward models that provide
-step-level guidance (+4--8 MATH points), and hybrid verification enabling reliable
-self-evaluation. The combination yields 83.4\% MATH, 97.9\% GSM8K, and 48.2\% AIME
-for Zen-32B. Open problems include proof-level verification for advanced mathematics
+step-level guidance (+4.6 to +4.8 MATH points), and hybrid verification enabling
+reliable self-evaluation. The combination yields 83.4\% MATH, 97.9\% GSM8K, and
+48.2\% AIME for Zen-32B with the PRM. Open problems include proof-level verification for advanced mathematics
 and step-level data collection at Olympiad difficulty.
 
 \bibliographystyle{plain}

zen/zen-multilingual.tex

@@ -23,7 +23,7 @@
 \maketitle
 
 \begin{abstract}
-We present Zen Multilingual, the multilingual capabilities of the Zen MoDE (Mixture of Distilled Experts) model family covering 110 languages including 48 low-resource languages with fewer than 1 million tokens of training data. Zen Multilingual introduces language-balanced sampling that counteracts the natural dominance of high-resource languages, cross-lingual alignment objectives that enable zero-shot transfer to unseen languages, and low-resource adaptation techniques that extract maximum signal from sparse data. On FLORES-200 translation, Zen Multilingual achieves 38.4 average BLEU across 200 language pairs. On XCOPA cross-lingual commonsense, we achieve 84.2\% average accuracy across 11 languages. On XNLI natural language inference, we achieve 81.8\% average across 15 languages.
+We present Zen Multilingual, the multilingual capabilities of the Zen MoDE (Mixture of Distilled Experts) model family covering 110 languages including 48 low-resource languages with fewer than 1 million tokens of training data. Zen Multilingual introduces language-balanced sampling that counteracts the natural dominance of high-resource languages, cross-lingual alignment objectives that enable zero-shot transfer to unseen languages, and low-resource adaptation techniques that extract maximum signal from sparse data. On FLORES-200 translation, Zen Multilingual achieves 38.4 average BLEU across 200 language pairs. On XCOPA cross-lingual commonsense, we achieve 85.2\% average accuracy across 11 languages with eight-shot prompting. On XNLI natural language inference, we achieve 83.3\% average across 15 languages.
 \end{abstract}
 
 \section{Introduction}
@@ -202,7 +202,7 @@ Turkish & 82.8 & 87.4 \\
 Vietnamese & 86.4 & 90.2 \\
 Yoruba & 68.4 & 74.8 \\
 \midrule
-\textbf{Average (11)} & \textbf{80.6} & \textbf{84.2} \\
+\textbf{Average (11)} & \textbf{80.6} & \textbf{85.2} \\
 \bottomrule
 \end{tabular}
 \end{table}
@@ -233,7 +233,7 @@ Turkish & 80.2 & 78.6 \\
 Urdu & 78.8 & 77.2 \\
 Vietnamese & 82.4 & 80.8 \\
 \midrule
-\textbf{Average} & \textbf{81.8} & \textbf{80.1} \\
+\textbf{Average} & \textbf{83.3} & \textbf{81.6} \\
 \bottomrule
 \end{tabular}
 \end{table}

zen/zen-nano_whitepaper.tex

@@ -154,7 +154,7 @@ HellaSwag & 59.5\% \\
 \textbf{Metric} & \textbf{Value} \\
 \midrule
 Inference Speed & 450 tokens/sec \\
-Memory Usage (INT4) & 2 GB \\
+Memory Usage (INT4) & 0.3 GB \\
 Energy Efficiency & 98\% reduction \\
 Latency (First Token) & 15 ms \\
 \bottomrule
@@ -258,7 +258,7 @@ The model has been developed with careful attention to:
 \midrule
 FP16 & 1.2 GB & RTX 3060 \\
 INT8 & 0.6 GB & GTX 1660 \\
-INT4 & 2 GB & Raspberry Pi 5 \\
+INT4 & 0.3 GB & Raspberry Pi 5 \\
 \bottomrule
 \end{tabular}
 \caption{Hardware Requirements by Precision}

zen/zen-next_whitepaper.tex

@@ -258,7 +258,7 @@ The model has been developed with careful attention to:
 \midrule
 FP16 & 160 GB & 2x A100 80GB \\
 INT8 & 80 GB & A100 80GB \\
-INT4 & 40 GB & RTX 4090 \\
+INT4 & 40 GB & 2x RTX 4090 \\
 \bottomrule
 \end{tabular}
 \caption{Hardware Requirements by Precision}

zen/zen-omni_whitepaper.tex

@@ -256,8 +256,8 @@ The model has been developed with careful attention to:
 \toprule
 \textbf{Precision} & \textbf{Memory} & \textbf{Recommended Hardware} \\
 \midrule
-FP16 & 60 GB & A100 40GB \\
-INT8 & 30 GB & RTX 4090 \\
+FP16 & 60 GB & A100 80GB \\
+INT8 & 30 GB & A100 40GB \\
 INT4 & 15 GB & M2 Max MacBook Pro \\
 \bottomrule
 \end{tabular}

zen/zen-scribe_whitepaper.tex

@@ -145,7 +145,7 @@ Multilingual ASR & 5.2\% \\
 \textbf{Metric} & \textbf{Value} \\
 \midrule
 Inference Speed & 380 tokens/sec \\
-Memory Usage (INT4) & 3 GB \\
+Memory Usage (INT4) & 0.8 GB \\
 Energy Efficiency & 96\% reduction \\
 Latency (First Token) & 20 ms \\
 \bottomrule
@@ -249,7 +249,7 @@ The model has been developed with careful attention to:
 \midrule
 FP16 & 3 GB & RTX 3060 \\
 INT8 & 1.5 GB & RTX 2060 \\
-INT4 & 3 GB & Intel NUC \\
+INT4 & 0.8 GB & Intel NUC \\
 \bottomrule
 \end{tabular}
 \caption{Hardware Requirements by Precision}

zen/zen-synthetic-data.tex

@@ -23,7 +23,7 @@
 \maketitle
 
 \begin{abstract}
-We present the Zen Synthetic Data framework, a scalable pipeline for generating high-quality training data for instruction following, reasoning, domain specialization, and alignment. Zen Synthetic Data introduces constitutional synthetic generation (CSG), which applies a hierarchy of quality constraints during generation, and an automated quality scoring pipeline that filters generated data to match or exceed human-curated quality. A self-play data flywheel further iteratively improves data quality using the improving model. On downstream evaluations, models trained with 4.8 million synthetic CSG examples improve by an average of 8.4 points on MT-Bench and 6.2 points on MMLU compared to training on equal volumes of web-scraped instruction data. Human raters score synthetic CSG data at 4.21/5 versus 3.84/5 for filtered web data.
+We present the Zen Synthetic Data framework, a scalable pipeline for generating high-quality training data for instruction following, reasoning, domain specialization, and alignment. Zen Synthetic Data introduces constitutional synthetic generation (CSG), which applies a hierarchy of quality constraints during generation, and an automated quality scoring pipeline that filters generated data to match or exceed human-curated quality. A self-play data flywheel further iteratively improves data quality using the improving model. On downstream evaluations, models trained with 4.8 million synthetic CSG examples improve by 0.88 points on MT-Bench (7.84 to 8.72) and 4.0 points on MMLU (82.4\% to 86.4\%) compared to training on equal volumes of web-scraped instruction data. Human raters score synthetic CSG data at 4.21/5 versus 3.84/5 for filtered web data.
 \end{abstract}
 
 \section{Introduction}

zen/zen3-embedding_whitepaper.tex

@@ -27,8 +27,8 @@
 Zen3-Embedding produces 7680-dimensional semantic representations optimized for DSO
 (Decentralized Semantic Optimization) retrieval, achieving strong multilingual embedding
 quality across the MTEB benchmark suite while maintaining BitDelta-compatible compression
-for efficient deployment. At 7B parameters, Zen3-Embedding achieves MTEB average 74.3\%,
-BEIR nDCG@10 57.8\%, MIRACL 81.2\%, and retrieval nDCG@10 0.812, surpassing prior
+for efficient deployment. At 7B parameters, Zen3-Embedding achieves MTEB average 67.4\%,
+BEIR nDCG@10 57.8\% and MIRACL 81.2\%, surpassing prior
 embedding models with substantially smaller parameter counts. The model supports
 100+ languages through a shared multilingual encoder with language-balanced contrastive
 training, and its 7680-dimensional output space aligns natively with the DSO vector
@@ -241,7 +241,7 @@ Classification     & 78.6 & 74.9 & 12 \\
 Clustering         & 51.2 & 47.4 & 11 \\
 Pair Classification & 86.9 & 83.1 & 3  \\
 Summarization      & 30.4 & 28.7 & 1  \\
-\textbf{MTEB Average} & \textbf{74.3} & \textbf{70.1} & 56 \\
+\textbf{MTEB Average} & \textbf{67.4} & \textbf{63.6} & 56 \\
 \bottomrule
 \end{tabular}
 \label{tab:mteb}
@@ -349,7 +349,7 @@ Zen3-Embedding is the first 7B-scale model to implement full MRL across 6 dimens
 \label{sec:conclusion}
 
 Zen3-Embedding delivers frontier retrieval quality at 7B scale with native DSO
-protocol compatibility, achieving MTEB 74.3\%, BEIR 57.8\%, and MIRACL 81.2\%.
+protocol compatibility, achieving MTEB 67.4\%, BEIR 57.8\%, and MIRACL 81.2\%.
 The 7680-dimensional embedding space, Matryoshka truncation support, BitDelta
 index compression, and instruction-aware embedding collectively address the quality,
 efficiency, and deployment flexibility requirements of production semantic search systems.

zen/zen3-nano_whitepaper.tex

@@ -240,7 +240,11 @@ Zen2-Nano    & 34.1 & 40.5 & 18.7 \\
 
 \begin{table}[H]
 \centering
-\caption{Inference throughput across hardware targets (4-bit quantized)}
+\caption{Inference throughput across hardware targets (4-bit quantized). Batch size
+is not recorded, and the ms/tok column is the reciprocal of the tokens/second column
+rather than a separate measurement. One decode stream reads the resident weights once
+per token, which on the 4090 row (950 MB, 1{,}008 GB/s) bounds a single stream at about
+1{,}060 tokens per second.}
 \begin{tabular}{lcccc}
 \toprule
 \textbf{Hardware} & \textbf{Precision} & \textbf{Tok/s} & \textbf{RAM (MB)} & \textbf{Latency (ms/tok)} \\

zen/zen4-mini_whitepaper.tex

@@ -345,7 +345,10 @@ MLX 4-bit       & M2 (8GB)         & 560  & 2.8 GB \\
 CPU (Q4\_K\_M)  & Intel i9-14900K  & 42   & 2.8 GB \\
 \bottomrule
 \end{tabular}
-\caption{Zen4-Mini Throughput Across Hardware Configurations}
+\caption{Zen4-Mini Throughput Across Hardware Configurations. Batch size is not
+recorded. A single decode stream reads all 2.8 GB of INT4 weights per token, so the
+4090's 1{,}008 GB/s bounds one stream at 360 tokens per second; the figures here are
+aggregate across concurrent requests.}
 \end{table}
 
 \subsection{Comparison to 14B Tier}

zen/zen_family_overview.tex

@@ -370,7 +370,7 @@ Zen-Artist & 16GB & 8GB & 4GB & RTX 3060 & RTX 3080 \\
 Zen-Omni & 60GB & 30GB & 15GB & RTX 4090 & A100 40GB \\
 Zen-Coder & 240GB & 120GB & 60GB & A100 80GB & 2x A100 \\
 Zen-Next & 160GB & 80GB & 40GB & 2x RTX 4090 & 2x A100 \\
-Zen-Designer & 220GB & 110GB & 55GB & A100 80GB & 2x A100 \\
+Zen-Designer & 470GB & 235GB & 118GB & 2x A100 80GB & 3x A100 \\
 Zen-Scribe & 3GB & 1.5GB & 0.8GB & Phone (4GB) & Any GPU \\
 \bottomrule
 \end{tabular}