Skip to content
K = 0.731
2026.10.17 · Sat

#推理

#推理

First-party

Direct announcements from model vendors, companies and research institutions on their official sites, blogs and accounts.

10/16Fri
  1. Hugging Face Blog83

    Reflection releases Beam, a 501B open model for coding and agentic work

    Reflection released Beam on Hugging Face, a 501B MoE model with 23B active parameters per token, a 1M-token context window, and an Apache 2.0 license. The text-only model targets coding and agentic work, with BF16, FP8, and NVFP4 checkpoints and day-0 support in transformers, vLLM, and SGLang. Hugging Face provides deployment and post-training recipes alongside its own task evaluations and demos.

    Why it matters: The deployment guidance distinguishes active computation from total weight memory, helping teams assess hardware requirements for running and adapting a sparse model with a long context window.

  2. Hugging Face Blog74

    Granite Switch 4.1 Combines Chat, Decisions, and Speech in One Endpoint

    IBM released Granite-Switch-4.1-8B and -30B, combining chat, a general-purpose Decision aLoRA, a specialized Guardian aLoRA, and Granite-Speech-5.0-Turbo-CTC transcription through one hosted endpoint. IBM says the combined model uses approximately the same GPU VRAM as its base model, with adapters reusing the base model’s KV cache. Users can also compose custom Switch models from Granite Libraries adapters; an Adapter SDK for training adapters in Mellea is planned.

    Why it matters: Shared KV caches and modular adapters offer a concrete approach to combining general and specialized decisions with transcription while keeping GPU memory use close to the base model.

  3. Microsoft Research79

    Microsoft Research presents Magnetar for proof-based general-purpose computing

    Microsoft Research presents Magnetar, a zero-knowledge virtual machine that proves RISC-V program execution using security based only on standard hash functions. It supports programs written in Rust, Go, or C without requiring developers to rewrite them as cryptographic circuits. Across 15 benchmark programs, Microsoft reports average speeds 10 to 39 times faster than the fastest existing comparable zkVMs on the same machine, with proofs verified in about 12 milliseconds.

    Why it matters: The benchmark comparisons connect proving speed with general-purpose execution and standard hash functions, clarifying the tradeoffs that have limited broader adoption of verifiable computing.

10/15Thu
  1. Mistral AI77

    Mistral AI releases Timestral, a family of time-series foundation models

    Mistral AI released Timestral in four sizes, from 5M to 209M parameters, for probabilistic, multivariate forecasting with past-only and known-future covariates. Mistral reports that its 209M model ranks first among pretrained, zero-shot models across five metrics on GIFT-Eval, TIME and fev-bench, using a single checkpoint without ensembling. All sizes support CPU inference, and Mistral published a GitHub repository for running Timestral on users’ data; domain adaptation and fine-tuning are available through Mistral Forge.

    Why it matters: The comparisons across model sizes and CPU latency provide a concrete basis for assessing forecasting accuracy against compute requirements in production deployments.

  2. Hugging Face Blog74

    Apertus 1.5 Release Notes Detail Multimodal Understanding, Thinking Mode, and Longer Context

    Project Apertus released Apertus 1.5 in 8B and 70B sizes, adding image and audio understanding, optional thinking, a 262,144-token context window, and improved instruction-following and tool use. The weights were released on Hugging Face on July 24, 2026 under Apache 2.0; the models generate only text, and spoken-language processing remains experimental. The team also publishes training data, recipes, intermediate checkpoints, and reproducible evaluation records.

    Why it matters: Published intermediate checkpoints, training mixtures, and evaluation settings provide a concrete basis for investigating how individual training stages change model behavior and for reproducing reported results.

  3. Google Research78

    Google releases Deep Review for analyzing complex technical documents

    Google Research has publicly released Deep Review, an agentic AI system for technical verification and feedback on research papers and other complex documents, including legal and financial materials. Evolving from the Paper Assistant Tool, which provided optional feedback on more than 16,000 manuscripts, it combines document segmentation, adaptive compute budgets, parallel verification, and global synthesis. Users worldwide receive two complimentary reviews per day, with higher limits and API integration available through a Gemini API key; uploaded documents are never used to train, fine-tune, or improve Google models.

    Why it matters: The review pipeline offers a transferable approach to checking long documents by combining section-specific compute budgets with full-document context and final consistency checks across sections.

10/14Wed
  1. Hugging Face Blog79

    Liquid AI releases LFM2.5-24B-A2B for fast, private on-device agents

    Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning model with 24B total parameters and 2B active parameters, for on-device agents that reason and call tools. Its context window expands from 32K to 131K tokens, and both the reasoning and Base versions are available on Hugging Face. Liquid AI reports 184 tokens/s on an Apple M5 Max and competitive tool-use results, while acknowledging that larger models retain advantages on knowledge-heavy and coding benchmarks.

    Why it matters: The benchmark comparison helps distinguish workloads suited to local deployment, with strong tool use and instruction following alongside weaker knowledge and coding results than some larger models.

  2. Google DeepMind94

    Google DeepMind launches Gemini 4 Argon for complex reasoning and agentic tasks

    Google DeepMind has launched Gemini 4 Argon globally through its API and selected Google products, with access depending on subscription or service tier. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off and a 1M unified input-output token limit through the Gemini API. Google reports improvements across coding, professional workflows, scientific research, and robotics, alongside stronger agent safeguards and prompt injection resistance.

    Why it matters: Google’s reported gains in code submissions and reductions in rollbacks connect Argon’s engineering capabilities to delivery speed and code stability in internal workflows.

  3. Google DeepMind82

    Google DeepMind Introduces Gemini 4 Argon for Robotics

    Google DeepMind introduces Gemini 4 Argon, a general-purpose physical reasoning model for robot control, simulation reconstruction, physical understanding, and self-improvement. Google DeepMind reports an 87% ERQA score using pure reasoning and a simulation-generated training example that raises real-world robot success from 56% to 75% while reducing completion time from 35s to 24s. Argon is available today to Google AI Ultra subscribers through Google AI Studio and Google Antigravity, and through the paid Gemini API tier using gemini-4-argon in the Interactions API.

    Why it matters: The simulation-to-hardware example provides a concrete basis for assessing how generated training episodes can improve robot task success and completion speed within a reported setup.

10/13Tue
  1. AWS Machine Learning Blog75

    Build long-horizon agents with Grok 4.7 on Amazon Bedrock

    AWS demonstrates how to build and evaluate a tool-calling agent using Grok 4.7 on Amazon Bedrock, comparing four reasoning effort levels across three synthetic tasks. The agent answered 67 of 72 runs correctly; xhigh achieved 18 of 18, while low achieved 15 of 18 using about an eighth of the output tokens. The guide covers task design, automated grading, production monitoring, and operational controls, noting that these small-scale results are indicative rather than statistically conclusive.

    Why it matters: Tasks with provably unique answers and repeated runs provide a reusable way to compare reasoning effort, while separate correctness and token measurements make the trade-off concrete.

10/7Wed
  1. Hugging Face Blog62

    Nemotron fine-tuning reaches gold-medal level at both IOI 2026 and IMO 2026

    The NVIDIA team started with Nemotron 3 and used supervised fine-tuning, reinforcement learning, and a generate-verify-improve reasoning process to build specialized models that reached gold-medal thresholds at both IOI 2026 and IMO 2026.

    Why it matters: The source gives a complete recipe for fine-tuning Nemotron 3 to gold-medal-level results at IOI and IMO, which readers can use to reproduce its data and reasoning workflow.

  2. OpenAI News62

    OpenAI announces progress by an internal frontier model on open math problems

    OpenAI released new results by an internal frontier model on open math problems and published Lean formalizations and research details on GitHub.

    Why it matters: OpenAI made results by an internal frontier model on open math problems public alongside Lean formalizations, showing how model reasoning can be combined with formal verification.

10/2Fri
  1. NVIDIA Blog78

    NVIDIA GPU-accelerated GPT-6 Astra Ultrafast launches

    GPT-6 Astra Ultrafast is now available in the OpenAI API and to eligible ChatGPT Work and Codex users. It runs on NVIDIA Blackwell GPUs.

    Why it matters: The source gives Astra Ultrafast's speedup over standard mode and access points, helping readers assess its practical effect on coding-agent loops.

10/1Thu
  1. Google DeepMind88

    Google DeepMind releases frontier model Gemini 4 Argon

    Google DeepMind released frontier model Gemini 4 Argon, initially making it available through the Fairwind Program to trusted cybersecurity defenders before gradually expanding access. Pricing starts at $2 per million input tokens and $10 per million output tokens.

    Why it matters: Gemini 4 Argon raises the maximum output to 1M tokens and reports multiple benchmark results and internal deployment data, offering a view of its limits on long-running workflows.

9/16Wed
  1. NVIDIA Blog62

    NVIDIA Vera Rubin NVL72 makes its MLPerf Inference v6.1 debut with leading performance

    NVIDIA submitted preview results for Vera Rubin NVL72 for the first time in MLPerf Inference v6.1, reporting up to 3.7× the throughput of GB300 NVL72 on Qwen3-VL and up to 2.5× on DeepSeek-R1.

    Why it matters: MLPerf Inference v6.1 reports throughput, scaling efficiency, and software-optimization gains for Vera Rubin NVL72 versus GB300 NVL72, offering a basis for judging inference infrastructure value.

  2. NVIDIA Blog72

    NVIDIA shares efficiency progress for Vera Rubin and DSX at AI Infra Summit

    At the AI Infra Summit, NVIDIA announced several advances in Vera Rubin and its DSX platform, shifting AI infrastructure measurement from peak performance to useful agentic tokens per megawatt.

    Why it matters: The article reports several measured token-output figures per megawatt for Vera Rubin and DSX, showing how efficiency metrics for AI factories are changing.

8/25Tue
  1. Hugging Face Blog62

    IBM releases Granite 4.2 reasoning models in 3B, 8B, and 30B sizes

    IBM's Granite team has released Granite 4.2, its first family of dense decoder-only reasoning models, in 3B, 8B, and 30B sizes. They were post-trained from the Granite-4.1 base and are all licensed under Apache 2.0.

    Why it matters: The article discloses Granite 4.2's architecture, SFT data mix, and multi-stage RL curriculum in detail, offering a look at concrete post-training methods for reasoning models.

  2. Hugging Face Blog60

    Multiverse Computing proposes Quantization-Aware Healing: 4-bit compressed model beats its bfloat16 original on seven of nine benchmarks

    Multiverse Computing has published a paper proposing Quantization-Aware Healing (QAH). After compressing GPT-OSS 120B to 60B parameters and quantizing it to MXFP4, the method distills directly from the original model before compression rather than from a recovered bfloat16 checkpoint.

    Why it matters: The paper presents a repair recipe that distills from the original model after compression and quantization, with stability comparisons against QAT.

8/12Wed
  1. Google Research60

    Google Research proposes knowledge profiling: retrieval, not encoding, is the bottleneck for frontier LLM factuality

    Google Research has proposed a knowledge-profiling framework that distinguishes factual errors caused by encoding failures from those caused by retrieval failures, and released WikiProfile, a benchmark containing 2,150 Wikipedia facts.

    Why it matters: By separating encoding failures from retrieval failures, knowledge profiling offers a new way to diagnose the sources of factual errors in frontier models.

7/29Wed
  1. Berkeley AI Research62

    From CUDA to MLX: how K-Search brings kernel-optimization experience to Apple Silicon

    Berkeley AI Research and IBM Research have extended the evolutionary kernel-search framework K-Search to MLX and added a structured CUDA-to-MLX translation layer that automatically adapts existing CUDA-kernel knowledge to Apple Silicon.

    Why it matters: The article details the CUDA-to-MLX translation layer and reports two kernel benchmarks, showing how kernel knowledge can be transferred across platforms.