#推理

NukeTorch: Four MacBook Pro M5 Max Macs Serve NVIDIA GR00T 2.45x Faster Than Fastest Published H100 Result

Fenon AI's NukeTorch engine ran NVIDIA's GR00T N1.7 robot policy on four MacBook Pro M5 Max machines at 152.07 action chunks per second with 192 clients, 2.45x the fastest published H100 figure and 4.24x NVIDIA's own published number. For a single request the H100 still wins at 16.08 ms versus 44.34 ms, and the four Macs list at $19,996 against at least $25,000 for one H100 card. The one-Mac version is a free download, while the multi-machine build requires an evaluation license.

Rethinking Memory for AI Inference

Moving compute closer to memory — via processing near memory (PNM) and processing in memory (PIM) — can cut data movement, power consumption and latency for AI inference at the edge. Cadence's Frank Ferro says LLMs moving into physical devices are forcing architects to rethink memory partitioning, while Quadric's Steve Roddy notes inference's known data access patterns suit ahead-of-time compilation. Software support, standardization and thermal management remain barriers to adoption.

Google releases TPU Sync, an open-source KV cache transfer library for Cloud TPUs

Google Cloud has released TPU Sync, an Apache-2.0 C++ library that moves KV caches and model weights across TPU HBM, pinned host DRAM, shared storage and the data center network. In internal July 2026 tests on TPU v7x, Google reports over 90% lower KV transfer overhead, with Qwen3 235B median KV transfer latency dropping from several seconds to about 230 ms and average TTFT falling 60% for 235B and 90% for 480B at low concurrency. The library integrates with PyTorch via TorchTPU and with JAX, serving as the KV connector in vLLM-TorchTPU and the P/D transfer backend in SGLang-TorchTPU, though the README says it is not yet recommended for general use.

Why I Love AI: A Researcher's Case for the Field That Succeeds by Failing

A computer science researcher who began studying AI twenty years ago explains why the field still earns a place in their bio. AI's reputation for failure stems from the "AI Effect," where working systems stop being called AI, leaving only unsolved problems under that label. The field is built on approximating answers rather than achieving perfect ones, as the Travelling Salesman Problem illustrates.

SingVL-4B GGUF Released: GGUF Build of Sing VL S1 4B Multimodal Decision Model on Qwen3.5

Sing VL S1, a 4B-parameter multimodal decision model built on Qwen3.5, now has a GGUF release for llama.cpp CPU inference. It ships in BF16 (8.42 GB), Q8_0 (4.48 GB) and Q4_K_M (2.71 GB) decoders plus a shared F16 vision projector, with a default 4096-token context. GGUF validation covers 9 examples and 11 questions; full public scores for the GGUF variants are still pending.

Google releases Gemini 4 Hydrogen, the lowest-cost model in the Gemini 4 family

Google released Gemini 4 Hydrogen, the lowest-cost model in the Gemini 4 family, priced at $1.50 per million input tokens, $0.075 for cached input and $7.50 per million output tokens. Simon Willison notes that Carbon and Argon both cost $2.00/$0.10/$10.00 until Nov 30, 2026 and double from Dec 1, 2026, so Hydrogen is currently only about 25% cheaper than its larger siblings, a gap that widens to roughly 75% after the increase. Hydrogen is also the new free tier model and will power AI features in Google Search.

Same event; the pick carries“Google DeepMind releases Gemini 4 Hydrogen, a faster Gemini 4 model for coding and agentic workflows”

InferSSD runs 800B-class LLMs on consumer hardware by keeping expert weights on NVMe

InferSSD, an open-source inference runtime built on llama.cpp, runs 763B-class DeepSeek-V4.1-Flash-1017 and 816B-class Muse-Spark-Open on consumer machines by keeping routed expert weights on NVMe storage while attention, shared layers and the KV cache stay in RAM or VRAM. On an RTX 5090 with 32 GB RAM and 32 GB VRAM it reports 164.5 tokens/s prefill and 7.27 tokens/s decode for DeepSeek at a 4K prompt tier, versus 8.06 and 0.867 tokens/s for llama.cpp on the same system. The article also estimates that heavy daily use could mean roughly 85 TB of host reads per day and lays out a conditional SSD wear model, noting the figures are assumptions to test rather than a promised drive lifespan.

Google releases Gemini 4 Hydrogen, a fast execution model for coding and agentic workflows

Google introduced Gemini 4 Hydrogen, a model built for rapid execution in everyday coding and agentic workflows, rolling out globally across the Gemini API, AI Studio, Gemini for business, the Gemini app, Google Search, and Google Antigravity.

Same event; the pick carries“Google DeepMind releases Gemini 4 Hydrogen, a faster Gemini 4 model for coding and agentic workflows”

Air-LLM: Broadcasting AI Weights to Edge Devices via RF Computing for On-the-Fly LLM Inference

Air-LLM is an edge LLM inference architecture in which a central radio such as a 5G base station broadcasts model weights over the air, and edge users complete GEMV directly in the RF domain using RF mixers, so devices never store or load weights. Tested on NVIDIA Sionna ray-traced channels in two real urban scenes with a real RF mixer profile, it limits WikiText-2 perplexity degradation on LLaMA-3.1-8B to 4.0% while saving 157.7x/40.4x energy versus FP16 and weight-only quantization baselines. With 20 users, airtime is 104.1x/26.0x shorter, and because the central radio stays user-unaware, one broadcast serves unlimited users in coverage.

Cloudflare Workers AI previews Clairic, a System One-style extraction model

Cloudflare Workers AI has released Clairic in preview, a strongly typed extraction model that pulls, computes, estimates and forecasts values from inputs and returns confidence scores and probabilities for each output. The model has a 65,536-token context window, supports vision, offers zero data retention and is priced at $0.10 per M input tokens. In the documented invoice example, Clairic extracts fields such as invoice number, due date and total, returns null with p_null 1 for a missing PO number, and gives a median of 30 with a p10-p90 interval of 21 to 38 for days_to_pay.

Trevor Paglen on AI mind-reading and the erosion of inner privacy

Artist Trevor Paglen recounts a two-day fMRI experiment at Cornell University where an AI system, NeuroGen, predicted his brain responses and generated images to elicit specific activity, leaving him feeling his mind had been read. He traces the field from Jack Gallant's 2011 blurry video reconstruction to Alex Huth's LLM-based text decoding of brain activity, and notes Meta paid $500 million to $1 billion for a wristband startup that translates brain signals into computer commands. Rafael Yuste, who made mice see phantom images with lasers, co-founded the Neurorights Foundation to build legal guardrails around brain data.

Edge0 open-sources Edge0-145B and Edge0-120B on-device models with inference framework

Edge0 open-sourced Edge0-145B and Edge0-120B along with its unified inference framework and deployment tools, claiming Edge0-145B is the first open-source 145B-parameter on-device model. The company reports 29GB peak memory, 1,100+ tokens/s prefill, 80+ tokens/s decode and roughly 120W typical power on a compatible laptop, with 128K–256K context. It cites internal scores of 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 90.4 on GPQA-Diamond and 80.6 on IFBench, and says weights are on Hugging Face with day-one support for macOS, Windows and Linux.

Google releases Gemini 4 Hydrogen, now the default model in AI Mode and AI Overviews

Google released Gemini 4 Hydrogen, a new model it says gives more helpful answers to complex questions, including questions with multiple parts, hard-to-understand typos, or highly specific requests. It is available starting today in Search, the Gemini app, all tiers of Google AI Studio, the Gemini API, Gemini for business, and Google Antigravity for AI Pro and Ultra subscribers, and Google Search VP Robby Stein says it is now the default model in AI Mode and is used for the hardest questions in AI Overviews.

Same event; the pick carries“Google DeepMind releases Gemini 4 Hydrogen, a faster Gemini 4 model for coding and agentic workflows”

Four Amazon Bedrock inference choices that affect cost

Amazon Bedrock's cache duration, service tier, reasoning effort, and maxTokens settings can change inference cost without altering the model or prompt. A 5-minute cache write on Claude Sonnet 5 and Claude Opus 5 costs 1.25x the input rate, a 1-hour write costs 2x, and cache reads cost 0.1x; Flex tier requests to unsupported models fail with a ValidationException.

Google rolls out Gemini 4 Hydrogen, replacing 3.8 Flash across its products

Google has released Gemini 4 Hydrogen, which replaces Gemini 3.8 Flash across the company's products with immediate access rather than the staged, paid-tier-only rollout used for its recent Argon and Carbon releases. Google says Hydrogen runs at one of four thinking levels (low, medium, high, and a new xhigh) and is 2.5 times faster than Flash 3.8 even in high mode, while cutting fabricated claims by more than 5x. It also more than triples 3.8 Flash's Terminal-Bench score and reaches 54 on the Artificial Analysis Intelligence Index in xhigh mode, versus 50 for Argon.

Same event; the pick carries“Google DeepMind releases Gemini 4 Hydrogen, a faster Gemini 4 model for coding and agentic workflows”

Google releases Gemini 4 Hydrogen to power more complex Search answers

Google released Gemini 4 Hydrogen, a fast-execution model it calls its biggest Search improvement yet, and is rolling the update out to all Search users globally. Google says Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash and beats the more powerful Gemini 4 Argon on the Artificial Analysis Intelligence Index at its default high thinking level. In Search it improves intent understanding, typo handling, fan-out subqueries and visual responses in AI Mode, and it is also coming to AI Studio, the Gemini API, the Gemini App and Antigravity.

Same event; the pick carries“Google DeepMind releases Gemini 4 Hydrogen, a faster Gemini 4 model for coding and agentic workflows”

Google DeepMind releases Gemini 4 Hydrogen, a faster Gemini 4 model for coding and agentic workflows

Google DeepMind introduced Gemini 4 Hydrogen, a Gemini 4 model built for fast everyday coding and agentic execution rather than the long-horizon reasoning of Gemini 4 Carbon and Argon. It scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level versus 50 for Argon, reaches 54 at a new xhigh setting, and is priced flat at $1.50 per million input tokens and $7.50 per million output tokens across its 1 million-token context window. Google also launched Orchestrator in Google Antigravity, which switches between models like Carbon and Hydrogen within a session and reached near-Carbon task success on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster.

OpenAI analysis traces how training choices make frontier models hide their reasoning

OpenAI's safety team analyzed models from GPT-5 through GPT-6 Astra and traced the erosion of readable chain-of-thought to four training factors: rising base-model capability, architecture choice, reinforcement learning after pretraining, and inheritance from teacher models during distillation. Silent reasoning grows with pretraining quality, one anonymized architecture showed a much steeper gain on multi-step planning without a scratchpad, and more RL training made Astra-class and internal Terra models better at steering their own reasoning logs. OpenAI says it needs new monitoring methods beyond reading the scratchpad, and notes that a smaller model reasoning out loud can match a larger one with a short chain of thought at higher compute cost.

OpenRouter compares self-hosted LLM and API costs with break-even load calculations

OpenRouter published a break-even analysis showing when self-hosting an open-weight LLM costs less than calling it through an API, using per-provider prices as of October 8, 2026. At those prices, Qwen3.8 27B on one rented RTX 3090 at $0.50 an hour breaks even at 6 to 12% load, while DeepSeek V4.1 Flash on four H200s never beats the cheapest zero-data-retention endpoint. The post also notes that provider.zdr and the EU and US domains cover most privacy needs without self-hosting, and recommends running a steady base load locally while sending peaks to the API.

vLLM adds day-0 support for Thinking Machines' Inkling 2 Small

vLLM supports Thinking Machines' Inkling 2 Small on day 0, a multimodal text-and-vision MoE with 594B total and 17B active parameters and up to 1M tokens of context. The integration adds kernel fusion with PDL, sequence parallelism for prefill that cuts an 8K-token prefill step by 22% at TP4 and 26% at TP8, and a Gluon relative attention decode kernel that drops a batch-1 decode step at 100K context from 15.1 ms to 4.8 ms. vLLM reports up to 630 TPS/user at TP4 on GB200 with acceptance length 3.5, and also supports LoRA and RL via SkyRL.

vLLM serves DeepSeek-V4.1-Flash on Vera Rubin NVL72 at 58x GB300 NVL72 throughput

vLLM reports serving DeepSeek-V4.1-Flash on a single NVIDIA Vera Rubin NVL72 rack at 1.7M tokens/s per GPU with 387 tokens/s per user, 58x the GB300 NVL72, peaking at 2.52M tokens/s and 1,810 tokens/s for a single user. The gains come from kernels rebuilt for Rubin's split memory domains and FP4 tensor cores, a specialized TP8 decode path, and rack-wide KV and Engram pools shared over NVLink. With prefixes pooled across the rack and Dynamo workload-aware routing, the prefix hit rate holds at 96.5-97% even at high concurrency.

Semantic diffing of formal policies with anti-unification

An open source engine computes semantic diffs of formal policies using anti-unification over an e-graph, showing shared structure once and keeping both versions of each differing part. It targets autoformalization outputs in SMT-LIB, Cedar, Dogwood, OpenShell policies and CEL, where an LLM writes the artifact but plays no part in computing the diff.

Google launches Gemini Decisions API with Gemini Decision 1 at $0.10 per 1M tokens

Google launched the Gemini Decisions API, powered by a new model called Gemini Decision 1 (gemini-decision-1), which does not generate text but scores up to 1,000 typed questions against one shared input and returns calibrated probabilities. It reads text, images, audio, video and PDFs in a single prefill pass, supports yes/no, choice and score question types, and is priced at $0.10 per 1M input tokens with zero output tokens billed. Google reports a 262,144-token context, 150-250 ms end-to-end latency, and a top-1 ECE of 0.0250, about half that of TypeSafe Jev 1.14, though it trails Jev on the When2Call benchmark (79.9% vs 82.0%) and costs more per token ($0.10 vs $0.042).

AWS re:Invent 2026 AI/ML Track: Attendee's Guide

AWS will hold its annual re:Invent conference from November 30 to December 4, with the AI/ML track offering more than 140 unique sessions out of over 2,000 total learning sessions. The guide maps sessions to the AWS AI stack, spanning AI chips (AWS Trainium and NVIDIA GPUs), models and inference on Amazon Bedrock and SageMaker AI, and agent development with Amazon Bedrock AgentCore. Sessions are organized from newcomer content to technical deep dives, and many can be viewed on demand.

Google DeepMind launches Gemini Decisions API and Gemini Decision 1 model

Google DeepMind introduced the Gemini Decisions API, built for closed-ended decisions such as yes/no questions, multiple choice, and scale-based ratings, with the underlying Gemini Decision 1 model processing up to 1,000 independent questions in a single pass across text, images, audio, video, and PDFs. It returns probabilities with confidence values in 150 to 250 milliseconds and costs $0.10 per million input tokens with no output-token charge, which Google says is about five times cheaper than Gemini 3.5 Flash-Lite. The API is available now through Google AI Studio, with integrations for OpenRouter, LangChain, and Vercel rolling out, and Google says it already uses it internally for model routing in Gemini Notebook, design prototyping in Stitch, and safety checks in Google Search.

Same event; the pick carries“Google launches Gemini Decisions API and Gemini Decision 1 model for fast closed-set decisions”

Google launches Gemini Decisions API and Gemini Decision 1 model for fast closed-set decisions

Google introduced the Gemini Decisions API, powered by the new Gemini Decision 1 model (gemini-decision-1) and available through the Gemini API in Google AI Studio. A single multimodal input of up to 256K tokens can be checked against up to 1,000 yes/no, multiple-choice and rubric-scoring questions in one forward pass, returning probabilities and confidence scores at 150-250 ms end-to-end latency for $0.10 per 1M input tokens with no output-token charges. Google says the model leads its benchmark suite with the lowest calibration error and is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API; internal uses include model routing in Gemini Notebook, wireframe generation in Stitch and adversarial-content checks for MCP tools in Search.

Anthropic makes Claude Opus 5.5 the default model on Pro, Max, Ultra and Team plans

Claude Opus 5.5 is now the default model in Claude on Pro, Max, Ultra and Team plans, and Anthropic recommends it as the daily driver for research, analysis and finished documents, decks and spreadsheets. Anthropic says Opus 5.5 at Medium effort matched or beat Opus 5 at High on its coding and knowledge-work evaluations, and in an internal merger-analysis test it finished in 63 minutes at half the cost of Opus 5 on the same task. The post also lists prompting changes for the new model, including dropping 'think hard' instructions, starting at Medium effort and attaching charts instead of retyping their values.

inclusionAI Releases LemonDecider 4B v3 Decision Model Based on Qwen3.5-4B

inclusionAI released LemonDecider 4B v3, a decision model fine-tuned from Qwen3.5-4B that returns choice, yes/no or ordinal probabilities via a single joint forward pass instead of generating answer tokens. It was trained on about 1.20M decision/QA rows and scored 61.67 on the public Decision Index 0.4 suite, with median request latency of 49.9 ms.

inclusionAI Releases SingVL S1 4B Multimodal Decision Model on Qwen3.5

inclusionAI released SingVL S1, a 4B multimodal decision model built on Qwen3.5 that answers named questions over text or JSON and up to four images in a single Transformers forward pass. It supports choice, yes/no (noul) and 2–10 level score questions, scoring 73.94 on the Public Vision Index and 69.76 on the Public Text Index. The model ships with Python inference code and a FastAPI HTTP server.

What Happens to Mathematics When AI Becomes an Oracle?

Applied mathematicians Steven Strogatz and Alex Townsend explore how "big math" is outgrowing human comprehension in their new book Big Math: The Hidden Codes That Run Our World. The book examines what it means as LLMs, AI chatbots and eventually quantum computing take on math beyond the capacity of finite human brains, and how society might adapt.

Redwood Research Questions Whether GDM Intentionally Trains Gemini 4 on Chain of Thought

Redwood Research says it is unclear whether Google DeepMind intentionally trains Gemini 4 on chain of thought, after the Gemini 4 Safety report cited an auxiliary reward that observes thoughts as context. GDM staff told Redwood they were not training on CoT but could not explain the technical details, while a GDM blog post says the model is never penalized for what it writes in its chain of thought. Redwood warns that CoT rewards biased against misaligned reasoning could suppress honest verbalization and reduce CoT faithfulness.

NaiveAI's Naive-N0.5-Flash claims 2,100 tokens per second: how well does it code?

Latent Space tested NaiveAI's open-weights Naive-N0.5-Flash, a coding model derived from Xiaomi's MiMo-V2.5 and paired with the NaiveRT inference engine, against DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash. In the author's runs it peaked near 2,000 tokens per second and sustained about 1,000 during thinking-heavy tasks, with a median response time of 3.1 seconds versus 5.3 and 10.1 seconds for the two comparison models. On hard HLE questions it averaged about 108,500 completion tokens with thinking on, roughly nine times GLM 5.3 Flash and 11 times DeepSeek V4.1 Flash 1017, and the author notes the sample was 10 questions per difficulty level with no accuracy grading.

Google DeepMind releases Gemini 4 Carbon for agentic coding, cybersecurity and scientific reasoning

Google DeepMind made Gemini 4 Carbon available to developers, enterprises, scientists and consumers, positioning it above Gemini 4 Argon for tasks that need deeper reasoning and more tokens. Carbon 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 priced at $2 per million input tokens and $10 per million output tokens, the same introductory price as Argon. Google also previewed /swarm in Google Antigravity, which turns a single prompt into a coordinated team of agents, and reported that Carbon blocked 99% of attack trajectories in simulated cyberattack testing.

OpenAI studies what affects CoT monitorability across the model lifecycle

OpenAI researchers examined how pretraining, architecture, reinforcement learning and model-generated data affect two proxies for chain-of-thought monitorability: no-CoT capability and CoT controllability. They report that no-CoT capability tracks lower base-model perplexity smoothly, that CoT controllability rises during RL even without process supervision beyond length constraints and final-answer grading, and that controllability can be passed to a student model through SFT on a teacher's rollouts. The authors caution that both proxies are imperfect and that models may become harder to monitor through their CoT as capabilities scale.

Meta AI launches Muse Ember, its smallest and most efficient model

Meta AI launched Muse Ember, described as its smallest and most efficient model, aimed at tool calls, coding subtasks, routing, extraction and long document parsing. For shorter prompts it is priced at $0.10 per million input tokens and $0.50 per million output, the same as GPT-6 Luna and Claude Haiku 5.5, while its long-prompt rates of $0.30 input and $1.50 output above 200k tokens place it between the two. Simon Willison notes the model sits on the Pareto line of Artificial Analysis's Intelligence Index vs. Cost per Task chart, and that its six reasoning levels produced pelican SVGs of comparable quality to Muse Wonder at 40x the price.

Anthropic releases Claude Fable 5.2 and Mythos 5.2, retaking the top LLM spot as its IPO roadshow begins

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, which top most independent leaderboards, including Artificial Analysis Intelligence Index v5 at 65 and SemiAnalysis GDPx at 16.4%. Pricing drops 20% to $8/$40 per 1M input/output tokens, but SemiAnalysis reports cost per GDPx task rose from $67.16 to $88.15 because the model writes hundreds to thousands of test cases and spawns subagents to check its work. Multiple evals show the max effort level underperforming high, and Fable is capped at 50% of a weekly subscription limit.

Hugging Face explains decision models and their support in Transformers and llama.cpp

Hugging Face published an explainer on decision models, which take a state plus typed questions and return probabilities over predefined answers instead of generating text. The post details the choice, score and noul primitives, the confidence formula, and new support through Transformers' DecisionPipeline and AutoModelForDecision, a shared /v1/systemone endpoint in Transformers and llama.cpp, a Hub tag for discovery, and fine-tuning with the Trainer API. It notes that the same endpoint also works with compatible autoregressive LLMs by scoring answer logits in a batched forward pass.

Weighted State Souping for Efficient Long-Context Retrieval

Apple researchers and Harvard collaborators propose weighted state souping, a method that learns query-dependent weights for recombining cached SSM state encodings, improving long-context performance on single- and multi-hop document QA, needle-in-a-haystack retrieval, variable tracking, and tool use. A sparse variant achieves near-constant encoding cost with respect to context length while preserving or sometimes improving performance. The paper was accepted at the PALM workshop at NeurIPS 2026.

Apple Study: Gating Mechanism Explains Why SSMs Memorize Instead of Learning In-Context

An Apple and Harvard study shows that the gating mechanism in State Space Models (SSMs) causes them to first learn an in-weights "memorization" solution, delaying or even preventing convergence to correct in-context learning. This occurs even when the architecture and memory capacity impose no fundamental limits. Gating, however, often improves generalization to long sequence lengths.

Ethan Mollick tests Claude Fable 5.2 across coding, business and research, calling it almost AGI

2 on tasks spanning hardware hacking, business planning, humanities course building, academic research and a 450-world 3D portal project, and judged it almost AGI under his definition of matching most human experts at most computer-based intellectual tasks.