#推理

Advancing science & engineering with Antigravity’s /teamwork

Google describes how multi-agent workflows in Antigravity's /teamwork, built on Gemini 4 models, produced results on open problems including a full proof of the 2013 Courtade–Kumar conjecture formalized in Lean 4, a machine-checked formalization of the Weil conjectures spanning about 1.8 million lines of Lean, and 95% top-1 accuracy on molecular structure determination from NMR data. The post also reports engineering gains such as a memory-safe Rust rewrite of the Zircon microkernel adding about 800,000 lines, a libgav1 Rust port that became 20-30% faster than the original C++, and fleet-wide memory savings surpassing initial estimates of 1 PiB. Google says the tools are available to external researchers and developers in Antigravity today, with 10,000 academic researchers offered six months of free access to its highest usage tier.

When One GPU Isn’t Enough: Serving GLM-5.3 Across Two NVIDIA DGX Stations with vLLM

Running a frontier-scale model locally requires keeping all its weights in GPU memory, and the largest models no longer fit on one machine. Two NVIDIA DGX Stations connected over ConnectX-8 RDMA serve one model through a single endpoint, with tensor parallelism splitting the model weights and computation across both GPUs. The NVIDIA GLM-5.3 NVFP4 model is described as a quantized version of Z.ai's mixture-of-experts model.

DeepSeek-V4.1-Flash on Vera Rubin NVL72: 58x Throughput vs. GB300 NVL72 at 387 tokens/s

On a single NVIDIA Vera Rubin NVL72 rack, vLLM serves DeepSeek-V4.1-Flash at 1.7M tokens/s per GPU with 387 tokens/s per user, 58x GB300 NVL72, reaching 2.52M at peak throughput and 1,810 tokens/s for a single user. The gains come from kernels tuned for Rubin's split memory domains and FP4 tensor cores, specialized low-latency decode kernels, and a rack-wide KV cache and Engram tables shared over NVLink. The post details four optimization groups and reports a 96.5–97% prefix hit rate at high concurrency with workload-aware routing.

Deploy frontier AI models on AWS Trainium with SGLang-Neuron

Amazon announced that SGLang-Neuron, released by RadixArk with Annapurna Labs, is publicly available through an AWS Deep Learning Container, letting users serve models and generate RL rollouts on Trainium via the same SGLang interface. The out-of-tree plugin runs on native PyTorch, is selected with SGLANG_PLATFORM=neuron, and supports PD disaggregation with KV-cache transfer via NIXL over EFA. The post includes a walkthrough on a Trainium3 EC2 instance using the pytorch-inference-sglang-neuronx container to serve openai/gpt-oss-120b.

Using Claude Opus 5.5 as your daily driver for knowledge work

Anthropic says Claude Opus 5.5 is now the default model on Pro, Max, Ultra, and Team plans and recommends it for daily knowledge work such as research, analysis, and producing memos, reports, spreadsheets, and decks, priced at $4/$20 per million input/output tokens. The post describes internal tests on research reports and a fictional merger analysis, plus guidance to start at Medium effort, drop "think hard" prompts, and attach charts rather than retyping them. It also walks through a demo project where Opus 5.5 tested four theories for slowing retail growth and proposed an alternative explanation involving new stores opening too close to existing ones.

Inkling 2 Small on vLLM: Day-0 Support

vLLM announces day-0 support for Thinking Machines' Inkling 2 Small, a multimodal text-and-vision MoE model with 594B total and 17B active parameters, up to 1M tokens of context, and a native DSpark speculator. The post details the architecture (48 layers, 384 experts with top-8 routing, block attention residuals, gated attention, a 23-layer ViT image encoder, and mixed NVFP4/MXFP8 expert precision) and vLLM's optimizations, including kernel fusion with PDL, fused MoE activation quantization, sequence parallelism for prefill, and a Gluon relative attention kernel, reaching up to 630 TPS/user at TP4 on GB200. It also covers LoRA and RL (via SkyRL) support, a serve command, agentic-workload benchmarks, and accuracy checks on NIAH up to 1M tokens, MMMU-Pro and HLE.

Four Amazon Bedrock inference choices that affect cost

The post explains how four request-level Amazon Bedrock settings—prompt cache duration, service tier, reasoning effort, and maxTokens—can change token processing and billing rates without altering the model or prompt. It reports tests showing that a 1-hour cache pays off from the second request per hour when gaps exceed 5 minutes, that Flex on Nova 2 Lite processed the same tokens at half the Standard rate in us-west-2, that lowering effort from high to low cut median output tokens by 169% on an agent test, and that a 155-token maxTokens fence avoided 1,893 output tokens in one test. It recommends confirming model support, changing one setting at a time, and verifying effects through response metadata, CloudWatch metrics, invocation logs, and billing data.

Semantic diffing of formal policies with anti-unification

The post describes an open source engine that computes semantic diffs of formal policies by anti-unification over an e-graph, using declared algebraic laws and domain rewrites so that shared structure appears once and each genuine disagreement is kept as a Variants node. Worked examples cover SMT-LIB-style rules, Cedar and Dogwood authorization policies, and OpenShell and CEL encodings, showing how presentational differences are absorbed while real bugs such as a weakened conjunction or a dropped guard remain visible. The engine builds from source with a pinned Rust toolchain and reads a single S-expression file.

Weighted State Souping for Efficient Long-Context Retrieval

The paper proposes weighted state souping, a method that learns query-dependent weights for composing cached state encodings from an SSM, addressing the degradation of state souping and vanilla SSMs on long, crowded contexts. The authors report substantial long-context improvements across single- and multi-hop document QA, needle-in-a-haystack retrieval, variable tracking, and tool use, plus a sparse variant with near-constant encoding cost relative to context length. They also present theoretical arguments on why souping works and when it may fail.

On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models

This paper argues that the gating mechanism in modern recurrent networks explains both the successes and failures of State Space Models (SSMs) on in-context learning and retrieval tasks: theory and experiments show gating causes SSMs to first learn an in-weights "memorization" solution while delaying or preventing convergence to a correct in-context learning solution, even when the architecture and memory capacity impose no fundamental limits. The authors also find that gating is often beneficial for improving generalization to long sequence lengths.

Doubling LLM token rates in ComfyUI

ComfyUI rewrote the hot path of local text generation for Qwen3.x LLMs, letting the same checkpoints on the same GPU generate tokens up to twice as fast with no change to model or output quality; on an RTX 5090, Qwen3.5 9B went from 106 to 219 tokens per second. The changes include a re-laid-out GEMV kernel, fused prologue operations, new attention and DeltaNet decode kernels, and an adaptive-tree MTP speculative decoding algorithm, all available in ComfyUI v0.45.0 with the Text Generate node's mtp option left on auto.

We don’t understand whether GDM intentionally trains on CoT

Redwood Research researchers note that the Gemini 4 Safety report appears to say GDM uses auxiliary rewards that observe chain-of-thought content, citing Kaufmann et al. on training on CoT, while a GDM blog post says they never penalize what the model writes in its CoT and GDM staff said they weren't training on CoT but couldn't explain details. The authors argue that "orthogonal" CoT rewards are a weak guarantee and that avoiding training against monitors is insufficient to preserve CoT integrity, and they present an appendix experiment in which five LLMs rated transcripts with hack acknowledgments removed 20-50 points higher than the originals. They conclude they suspect a misunderstanding but want clarification and greater transparency.

Gemini 4 Carbon is now available

Google released Gemini 4 Carbon to developers, enterprises, scientists, and consumers, describing it as its strongest model for cybersecurity capabilities and long-horizon agentic workflows in knowledge work, software engineering, and scientific reasoning. It ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench at more than 2X lower cost per task than the next best external model, and is available via the API in Google AI Studio and Gemini for business at $2 per million input tokens and $10 per million output tokens. Google also previewed /swarm in Google Antigravity and reported safety results including a 0% prompt injection attack success rate on the Gray Swan IPI benchmark.

Running AV Reasoning Models Within Onboard Compute and Latency Limits

AV developers need reasoning models that fit onboard compute and latency constraints; larger models provide a foundation, but deploying them in a vehicle requires reducing model size, preserving useful reasoning, and validating performance on target hardware. The source mentions NVIDIA Alpamayo 2 Edge in this context, but the supplied material is truncated before detailing it.

How to make flash models even faster: optimizing DeepSeek V4.1 Flash 1017

Baseten describes using an internal harness to optimize 80+ kernels and bottlenecks for DeepSeek V4.1 Flash 1017, producing a Fast variant it reports at 991 output tokens/s on Artificial Analysis with the lowest time per task. The post details three hardware-level techniques: kernel fusion (including an improved fused router kernel cutting decode time by 2.7ms and a 30% faster mHC megakernel), Programmatic Dependent Launch to overlap kernel execution, and host loop minimization via breakable CUDA graphs in SGLang.

Meta's Muse Ember is the cheapest model scoring 36 on the Intelligence Index

Meta's Muse Ember scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 (max) at $0.14 per task versus $0.41, making it the cheapest model scoring 36 or higher. It stands out on AutomationBench-AA at 68.1% and scores 46 on the Coding Agent Index in Muse Code, but its AA-Omniscience accuracy of 21.1% is the lowest among similarly intelligent models, largely because it declines to answer 64% of questions, keeping its hallucination rate at 17.4%. Pricing is $0.10 input / $0.004 cached input / $0.50 output per 1M tokens, available via Meta's first-party API and Muse Code.

Your Experts Have Partners. There's HOPE for Keeping Them Together

This post introduces HOPE (Higher-Order Pruning of Experts), a second-order expert-pruning method for Mixture-of-Experts language models that accounts for pairwise expert interactions via an F-matrix and a quadratic program, rather than scoring each expert independently. Evaluated on Qwen3.5-122B, Qwen3.5-35B and GLM-4.5-Air with two calibration datasets, HOPE achieved the best average rank across conditions and gains of up to +6.1% on agentic coding tasks over the next-best method, with its advantage growing at aggressive pruning rates such as 50%. The post also describes a three-step calibrate/solve/prune package and notes limitations, including weaker advantages at 10–20% pruning and no cross-layer interaction modeling.

AI-assisted learning platform helps robots reason beyond literal commands

MIT and NYU researchers introduced FLEX, a flexible feature-editing pipeline for robot motion-planning that lets robots interpret natural-language commands and generalize beyond their literal meaning. FLEX uses a three-step process — Add, Generalize (via the LLM GPT-4o), and Mask — so a robot told "Don't spill on the laptop!" also avoids other electronics, and it was demonstrated in simulations of serving coffee, moving a candle, and carrying a bowling ball. The study, led by MIT assistant professor Andreea Bobu, will be presented at the annual Conference on Robot Learning.

SageMaker Inference now supports Multi-Instance GPU to maximize GPU utilization for model endpoints

Amazon SageMaker Inference now supports NVIDIA Multi-Instance GPU (MIG) partitioning on inference component endpoints, letting you divide a physical GPU into up to seven hardware-isolated partitions and place a different model in each. Partitioning is declared via AcceleratorPartitionLayouts on the endpoint config and AcceleratorPartitionType on each inference component, with no new API operations or IAM actions. In testing, a mig-3g.40gb partition of an H100 delivered 53–89% more output tokens per second and 18–34% lower cost per million output tokens than a full A100-40GB above a workload-dependent concurrency crossover.

Zero Mismatch, Simple Async RL: REINFORCE Zero, DeepSeek-V4.1 and Kimi K3.1

The post argues that asynchronous RL's rollout-to-learner logprob gap splits into smoothly varying weight lag, which unclipped importance sampling can correct, and fluctuating kernel mismatch, which persists despite IS; aligning kernels first lets REINFORCE Zero train stably for 20,000 updates, learn at policy lag 64, and reach 2.40× the training-step throughput of synchronous training with aligned kernels. It reports zero mismatch (bit-for-bit logprob agreement) for DeepSeek-V4.1-Flash and Kimi K3.1 on day 0, and for DeepSeek-V4-Flash and Kimi K3.1 on AMD MI355X. Code and recipes are in Inferact/zero-mismatch-rl, with the paper at hijkzzz/REINFORCE-Zero and training code in NVIDIA-NeMo/labs-molt.

Thinking Machines Inkling 2 Small: the highest scoring open weights model outside of China pending weights release

Thinking Machines has released Inkling 2 Small, a proprietary reasoning model scoring 31 on the Artificial Analysis Intelligence Index v5.0.1, with Apache 2.0 weights planned for a later, unconfirmed date. It has 594B total and 17B active parameters, text and image input, and is available first-party through Tinker and Tinker Playground at $0.30/$1.00 per 1M input/output tokens. The model gains 18 points over Inkling Small, matches Gemini 3.8 Flash (high), and costs $1.18 per Intelligence Index task.

More Performance From Every Byte: How Full-Stack Codesign Solves AI’s Memory Challenge

NVIDIA describes how its full-stack codesign approach addresses growing GPU memory pressure from larger models, longer contexts and agentic applications by designing GPUs, CPUs, networking, storage and software together across a memory hierarchy. Techniques include NVFP4 for model weights and KV cache, distributing weights across up to 72 GPUs in NVL72, offloading inactive context to CMX context memory storage via Dynamo, and keeping RAG indexes outside HBM with Nemotron Retriever and cuVS. NVIDIA reports that from Blackwell Ultra to Vera Rubin, HBM capacity per GPU rose roughly 3% while agentic throughput per dollar rose up to 30x on SemiAnalysis AgentX.

vLLM Delivers Day-0 Support for Kimi K3.1 Pro

vLLM announces day-0 support for Kimi K3.1 Pro, Moonshot AI's 2.8T-parameter (104B active) multimodal MoE with a 1M-token context that adds a DSA sparse-attention indexer to all 24 full-attention layers of its KDA + MLA hybrid. The post details a packed KV cache for three kinds of state, token-sharded prefill attention (up to 1.69× faster TTFT at 340K tokens on 8× GB300), decode context parallelism with an exact top-k merge, a fused low-latency decode path, a DSpark speculator preview, AgentX benchmark results on GB300 and Rubin, zero-mismatch RL, and support on NVIDIA, AMD, TPU and Alibaba T-HEAD Zhenwu PPUs.

Together Compiler: Continual Compilation for AI inference

Together AI introduces Together Compiler (TKCC), a continual optimization loop that uses high-throughput agents to generate inference-engine optimizations, built on "persistent engines" that hot-reload candidate forward passes without restarts and a two-tier engine design balancing measurement throughput against fidelity. TKCC reports up to 2.6x higher TPS across frontier open-source models in under 24 hours, state-of-the-art low-latency results on SemiAnalysis' InferenceX benchmark (1111 TPS on MiniMax M3.1-Flash, 1205 TPS on DeepSeek v4.1-Flash), and 1.7x over open-source megakernels on Llama 1B at batch size 1, enabled by 252x higher experiment throughput than baseline agentic systems. The post also says TKCC generalizes to weather modeling, PDE, time-series, and biology workloads, and is serving trillions of tokens today.

Together Compiler: Continuous Compilation for AI inference

Together AI introduces Together Compiler (TKCC), a continual agent-driven system that recompiles inference engines as traffic drifts, using persistent engines, a two-tier fast/slow engine setup for throughput-fidelity balance, and a custom cost model for parallelism selection. TKCC reports 1.3-3x higher TPS across frontier open-source models in under 24 hours, state-of-the-art low-latency results of 1111 TPS on MiniMax M3.1-Flash and 1205 TPS on DeepSeek v4.1-Flash on SemiAnalysis' InferenceX benchmark, and 1.7x over open-source megakernels on Llama 1B batch size 1, enabled by 252x higher experiment throughput than baseline agentic systems. It also generalizes to weather modeling, PDE, time-series, and biology workloads.

Featured coverage of this story: 《Together Compiler: Continual Compilation for AI inference》

Fireworks unveils new products at Forge to help companies build specialized intelligence

At its first annual Forge conference, Fireworks launched products aimed at helping companies build AI around their own data and tasks, including Ultra-fast inference on NVIDIA Blackwell GPUs reaching up to 1,500 tokens per second, a half-price Background Mode, and multi-day prompt caching. It also introduced Fireworks Nexus for bringing open models into existing coding and coworking tools, No Eval and Specialized Model Inference for training and serving custom models, and Fireworks Sovereign, now in private preview, which runs its inference engine inside customers' own cloud or data centers. Fireworks says it already manages over 25,000 GPUs in customer environments.

Cost per task: Using cheap classifiers to route across Gemini models and thinking levels

The article argues that cost per successful task, not price per token, governs production economics for coding agents, and reports a 50-question auto-graded benchmark routing tasks with TypeSafe's Jev classifier across Gemini configurations. A frozen regex and prompt-length heuristic scored 93.8% on 16 development questions but fell 23.2 points to 70.6% on 34 held-out tasks, while Jev agreed with hand labels on all 34 and 48 of 50 overall; capping maxOutputTokens at 300 saved nothing on easy tasks and broke 7 of 25 hard ones. A router sending tasks to Flash-Lite first and escalating Jev-flagged complex ones to Flash on low thinking solved all 50, as did Flash on low thinking alone, and the article notes caching favors keeping long context-heavy turns on one model.

StateBench: Testing how language models track state changes over long documents

Ai2 introduces StateBench, a long-context state-tracking benchmark with four difficulty levels (r-trivial, aperiodic, soluble, and general finite) grounded in Krohn–Rhodes theory. Testing open pretrained models and frontier models including Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol and Luna, they found accuracy fell sharply on soluble and general finite problems as inputs grew longer, dropping to 13% or lower on the longest general finite problems. In from-scratch training of 275-million-parameter models, Gated DeltaNet and hybrid models generalized almost perfectly on the first three levels out to tens of thousands of assignments, while transformers never learned the hardest level.

Introducing Sarvam Prism

Sarvam released Sarvam Prism, a System One decision model built as a LoRA fine-tune of Sarvam 105B that returns a probability for each option in a single forward pass. Sarvam reports Prism scores 71.8 averaged over three indices at a 44 ms median response time on one B200 GPU, the highest-scoring model under 100 ms among the 25 it tested, and that its errors stay within its confidence bounds in English and Indian languages. Weights are on Hugging Face under Apache 2.0, and the model is also available on the Sarvam API via Sarvam Indus.

Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live iOS Apps

The authors propose an online reinforcement learning framework that trains UI navigation agents directly on live iOS applications, splitting the agent into a Planning agent and a UI Grounding model and using an LLM-as-a-Judge for dense step-level plus terminal rewards, optimized with PPO and step-level advantage estimation. They introduce GiGPPO, a neighbor-conditioned value estimation method that improves critic stability using reward signals from visually similar states across concurrent trajectories, backed by infrastructure orchestrating hundreds of concurrent virtual iOS devices. On UINavBench, iOSWorld, and AndroidWorld, their 8B RL agents substantially outperform SFT and terminal-only RL baselines, achieve nearly a sevenfold improvement over the SFT baseline in out-of-distribution generalization to Android, and outperform Gemini-3.5-Flash while tying Gemini-3.1-Pro on real iOS applications.

Decisions for a penny: introducing joey-1, our fastest decision model

Base Labs introduces joey-1, a decision model with 3B or fewer parameters that returns a label and calibrated probability score in a single forward pass, serving an average 2,000-token request at a P50 end-to-end latency of 20 ms (about 2.5x faster than elephant-1) at $0.01 per million tokens. Post-trained from Qwen3.5-2B using the same data and recipe as elephant-1 and mammoth-1, it is available on Baseten Model APIs with open weights on Hugging Face, and supports a 16k context window on Model APIs versus the full 262k on dedicated deployments.

Speeding up language models for travel applications

Researchers combined knowledge distillation, which trains a compact student model on a large teacher's outputs, with Medusa, a framework that predicts multiple tokens in parallel, and applied the pipeline with Booking.com to three travel features. The final model matched the quality of a 70-billion-parameter teacher while running 10 to 20 times as fast, with reported gains including 2.9% higher engagement versus GPT-3.5 for AI Trip Planner and 15-times-faster Smart Filters responses for 99% of queries. The work received the Industry Track best-paper award at ACL.

Welcome Le Chonk: Mistral Large 4 on Hugging Face

Mistral Large 4 ("Le Chonk") is a multimodal, multilingual Mixture-of-Experts model with about 1T total and 49B active parameters and a 1M-token context window, released in FP8 and NVFP4 variants under the Mistral Ecosystem Open-Weight License. It combines a 72-layer text decoder using MLA with a DeepSeek Sparse Attention indexer and a 1.6B Pixtral vision encoder, and the post reports 82% on the Artificial Analysis Cyber Index vulnerability reproduce-and-patch test and 93% on Cybench. Day-0 support covers Transformers, vLLM, llama.cpp, TRL fine-tuning, Inference Endpoints and Dell Enterprise Hub, with GGUFs from Unsloth and the llama.cpp team.

Day 0 support for Mistral Large 4 on vLLM

vLLM announces day-0 support for Mistral Large 4, a 1T-parameter MoE multimodal model with 49B active parameters per token, running on the V2 model runner with breakable CUDA graphs and a new vllm/models/mistral/ package. The work adds cross-layer KV sharing, DCP/PCP context parallelism, DSpark speculative decoding, NIXL-based disaggregated serving and HiSparse KV offload, plus day-0 ROCm support on AMD Instinct MI350 Series GPUs with MXFP4 experts. Support ships in the vllm/vllm-openai:mistral-large-4 Docker image, with FP8 and NVFP4 checkpoints and serving commands documented.

Le Chonk: Built for Coders, Trusted by Defenders.

Mistral released the open weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter natively multimodal model trained from scratch in Europe, reporting leading open-weight results on coding benchmarks such as FrontierCode 2.0 (19.7%) and Terminal-Bench 4.0 (41%), plus 51% on the Artificial Analysis Cyber Index. It offers three reasoning modes (Low, Medium, High) and is priced at $0.70 input / $0.07 cached input / $2.10 output per million tokens under a modified MIT license that requires a separate commercial license for Model-as-a-Service offerings.

Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5

Google's Gemini 4 Carbon tops the Artificial Analysis Intelligence Index v5 with a score of 59, leading 4 of 9 evaluations, and ranks second to Claude Opus 5.5 on the Coding Agent Index with 59 points. It scores 1609 Elo on AA-Briefcase (up 126 over Gemini 4 Argon) and 54 on AA-Omniscience v1.1, at $7.76 per Intelligence Index task on standard $4/$20 pricing. The model is being rolled out to selected users and is not yet publicly available, with a 50% launch discount and no confirmed end date.

Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design

Meta describes the next generation of its Andromeda ads retrieval engine, which uses a retrieval mixture-of-experts architecture and a single fused GPU pipeline to score every eligible ad at over one billion ads per second, a 100x throughput lift over the 2024 system. The system delivers near ranking-level model performance at retrieval scale with roughly 300x fewer inference-time FLOPs than the 2024 model, and Meta reports a 2.4% increase in conversions on Facebook since 2025. A joint collaboration with NVIDIA pushed individual components toward hardware limits, with the gating kernel reaching over 10 billion ads per second on a single GH200 and index matching exceeding 100 billion ads per second.

Announcing Artificial Analysis Intelligence Index v5

Artificial Analysis released Intelligence Index v5, its biggest upgrade since v4, adding FrontierCode 2.0 and Terminal-Bench-Science and upgrading HLE-Diamond, AA-Omniscience v1.1 and CritPt v2, with 60% of the Index now based on private test sets or held-out answers. Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55), while the open weights gap widens from 3 points to 12. It also launched the Progress Index v1, which tracks model capability over time on one continuous scale.

3 Ways to Boost MoE Inference on NVIDIA DGX Spark

Profiling a batch-size-1 evaluation of Inclusion AI's Ling-3.0-flash on NVIDIA DGX Spark showed decoding was memory-bound, with model weight reads driving execution time. The MoE computation and the output-layer projection that produces scores across the model stood out as the two key operations.

How Specializing the Vocabulary of SLMs Can Cut Inference Cost by 77% in Real Applications

Dharma-AI describes VocabX, a method that rebuilds a small model's tokenizer around the cost of each token rather than its frequency, combining sub-word merges with multi-word "macro" tokens under a shared cost criterion. In a production Portuguese essay grader built on Qwen3-4B, the approach cut average response length from 350 to 101 tokens and measured inference cost by 77.0% with the same base model and training recipe, while the bench score rose from 0.703 to 0.810. The article previews results from a VocabX paper still under review, noting the numbers may change before publication.