#Reasoning

Vercel Developers Can Use Perplexity Decider on AI Gateway

Vercel developers can now use Perplexity Decider on AI Gateway! https://twitter.com/vercel_dev/status/2111211692030152951

Perplexity Decider v1.2 and Decider Flash v1.2 arrive on OpenRouter

Aravind Srinivas calls OpenRouter the best place to use Perplexity Decider models, quoting OpenRouter's announcement. https://twitter.com/openrouter/status/2111200007437402243 OpenRouter says the open-weights models accept text or images and return typed answers with probabilities. Decider v1.2 is a new 27B checkpoint succeeding v1.1, while Decider Flash v1.2 is 4B, faster, and half the cost. http://openrouter.ai/perplexity/pplx-decider-v1.2 http://openrouter.ai/perplexity/pplx-decider-flash-v1.2

We just open-sourced another pair of state-of-the-art Decision Models:

Aravind Srinivas announced two open-sourced Decision Models: pplx-decider-flash-v1.2: 4B params, 1 cent per million tokens, and the top 4B model on the HF Decision Index. pplx-decider-v1.2: 27B params, 2 cents per million tokens, and #2 overall on the HF Decision Index.

Anthropic’s STEM Fellows Program explores more efficient quantum circuits with Claude

Anthropic reports that Claude Mythos 5 designed quantum circuits from plain-language instructions through its STEM Fellows Program. The circuits improve on the best published ones, sometimes by an order of magnitude or more.

Google releases Gemini 4 Argon with temporary half-price API access

Google is rolling out Gemini 4 Argon globally to developers, enterprises, and paid subscribers, emphasizing coding, professional knowledge work, multimodal understanding, and sustained task execution. API pricing is $2 per million input tokens and $10 per million output tokens through November 30, rising to $4 and $20 on December 1. Google reports that Argon worked unattended for 11 hours on ADK optimizations, with engineers merging two changes; Ethan Mollick described improvements but noted remaining weaknesses in complex multi-agent tasks.

How VLA and World Models Give Robots an Embodied Brain

The article describes how VLA models and world models are being developed and deployed in Shanghai: VLA enables robots to perform tasks, while world models predict what may happen before an action. Ant Lingbo’s “Lingbo General Brain” has been adapted to 17 robot brands and 20 configurations, and robots using it are working in multiple Shanghai pharmacy locations. The article says the two approaches are beginning to converge, while planning, standards, and evaluation are advancing.

Sam Altman welcomes Nick Cammarata’s optimistic outlook on mechanistic interpretability

Sam Altman says he is very happy to see Nick Cammarata’s post on interpretability. Cammarata says early work at OpenAI left him pessimistic, with primitive tools making about 2% of neurons legible. He now claims modern agentic tools make the vast majority legible in models up to trillions of parameters, and believes ambitious interpretability could be solved quite soon as automated scientists improve. https://already-clean.goodfire.com/

Elorian releases Glimpse, its first AI model built around visual reasoning

Elorian released Glimpse on Friday as a research preview with API access and a waitlist-only chat interface. Its visual scratch space lets it crop image regions, retrieve video frames, and run code to measure or compare visual content during reasoning. Pricing is $0.60 per million input tokens and $3 per million output tokens; Elorian’s evaluations report performance above GPT-6 Astra on scientific chart reasoning and short-form video captioning.

Google's Gemini 4 Argon is closer to the AI frontier than any Google model in months

Google DeepMind’s Gemini 4 Argon scored 97.5 percent on ARC-AGI-3 using the Provider Adapter Harness at a cost of $4,800, behind GPT-6 Astra’s 99.9 percent. Its score fell to 49.8 percent at low reasoning in the Standard Harness, where API problems prevented completion of all games. Epoch AI placed Argon at 165 ECI points, close to Opus 5.5 and GPT-6 Astra, although early estimates ranked it eighth in math and sixth in software engineering.

Perplexity adds Gemini 4 Argon, citing a new WANDR benchmark high

Aravind Srinivas says Gemini 4 Argon has set a new SoTA on the Wide-And-Deep-Research (WANDR) benchmark and is now available to all Perplexity users in fast search and Computer (agent) modes. https://t.co/0U671BhbMj The quoted @perplexity_ai announcement specifies availability for Pro and Max subscribers. https://t.co/n1TyI9RhBX

Bao Yu Shows How Handwired Turns Code Into an Exact Neural Network

Bao Yu says they made a tutorial video about turning code into a neural network without training, with every number matching. The post also links to English and Chinese versions of the video and says they tried Handwired by making examples such as countdowns and Snake.

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.

Is AI interpretability close to a breakthrough? A new essay has researchers divided

Cammarata argues that better tools and automated agents make neural networks more interpretable than earlier research suggested, while Google DeepMind's Neel Nanda disputes that a broad breakthrough is near. Cammarata reports preliminary findings that about 90 percent of neuron poles in Qwen3-32B have one main job and reinterprets classic superposition examples as related features. Nanda sees practical value in neuron analysis but expects significant gaps and maintains that superposition is real. The technical companion paper has not yet been released.

Reflection-Beam-1.0 gets Day-0 support in vLLM

Reflection-Beam-1.0 has Day-0 support in vLLM, with 501B total parameters, 23B active per token, and a 1M-token context window. Beam combines hybrid attention with Harmony-format reasoning and tool use, and has Day-0 RL support in SkyRL through LoRA, BF16 full fine-tuning and MXFP8 RL, with vLLM rollouts. vLLM worked with Reflection and Inferact on the serving path, from fused kernels to DFlash.

Reflection launches Beam, a 501B open-weight MoE model

Reflection launched Beam, a 501B MoE model, with open weights under the Apache 2.0 license and 40+ ecosystem partners. NVIDIA AI says it was trained on NVIDIA GB300 NVL72 and delivers over 70K tokens/sec/GPU on Blackwell Ultra out of the box on day 0. Reflection’s quoted announcement includes BF16, FP8 and NVFP4 weights, a DFlash2 drafter, and Mirror CLI; the Beam 1.0 NVFP4 checkpoint is available at https://huggingface.co/reflection/Beam-1.0-501B-A23B-NVFP4.

Liquid AI releases d1-ultra, leading the Artificial Analysis Decision Index v1.0

Liquid AI has released d1-ultra, which Artificial Analysis reports leads its Decision Index v1.0 with a score of 65. On AA-Omniscience-Decision, its non-hallucination rate rose to 85% from d1’s 2%, while accuracy fell to 32% from 41%. This non-hallucination metric measures how often the model says it does not know among questions it does not answer correctly. d1-ultra takes 92 ms per task and costs $0.042 per 1M input tokens, billed only on input tokens.

This paper shows why I am skeptical of the idea that frontier models can be easily replaced by cheaper ones (and why you should use the best models as orchestrators): Small improvements in accuracy & judgement compound over long-horizon tasks, creating huge gains in outcomes.

This paper shows why I am skeptical of the idea that frontier models can be easily replaced by cheaper ones (and why you should use the best models as orchestrators): Small improvements in accuracy & judgement compound over long-horizon tasks, creating huge gains in outcomes.

NVIDIA Vera Rubin NVL72 posts initial AA-AgentPerf results against GB300 NVL72

Artificial Analysis reports that NVIDIA's Vera Rubin NVL72 delivers up to ~5.7x GB300 NVL72's throughput per megawatt on DeepSeek V4 Pro in initial AA-AgentPerf results. Its advantage is largest at faster per-user speeds; peak-throughput gains range from 1.14x to 1.45x across the four tested models. NVIDIA submitted all Vera Rubin results using TensorRT-LLM and vLLM with NVIDIA Dynamo; the benchmark replays real agentic coding sessions and measures throughput per megawatt of measured GPU power alongside per-agent output speed.

Meta releases Muse Spark Open, scoring 39 on Artificial Analysis Intelligence Index

Meta has released Muse Spark Open, an 816B (42B active) open-weights model that Artificial Analysis reports scores 39 on its Intelligence Index, making it the #1 US open-weights model. Muse Spark Open (xhigh) scores 22 points above Muse Glimmer (high, 17), Meta's 30B dense open-weights model. https://t.co/jf31mQNINe

vLLM Reports First TPUv7 Submission to AgentX, Served with vLLM

We’re excited to see the first TPUv7 submission to AgentX, served with vLLM. Thanks to the @Google TPU, @RedHat_AI, and @inferact folks who put in the work to get it running, and we’re looking forward to more models landing on TPU through AgentX. https://t.co/Ea730oFhrd

Ethan Mollick Says Harnesses and Smart Models Could End Famous Hard Benchmarks

Harnesses + smart models = the end of the last famous hard benchmarks. Models can now work for extremely long times (bye, METR Long Tasks), they can output professional documents with correct math (so long, GDPval) and they can solve complex novel puzzles (farewell, ARC-AGI-3). https://twitter.com/arcprize/status/2110768288523251762

NVIDIA Nemotron 3.5 Super VL is now available on Ollama

NVIDIA Nemotron 3.5 Super VL is now available on Ollama. The model combines reasoning and visual understanding across a range of agent workflows. Ollama says it delivers up to 5x higher throughput on a single GPU for agentic tasks compared with other open models of its size.

NVIDIA introduces Nemotron 3.5 Bolt 24GB for local agents

NVIDIA AI introduces Nemotron 3.5 Bolt 24GB, saying it fits into 24GB of memory. The quoted @NVIDIARTXSpark post describes 74B-A9B parameters packed into 24GB of memory, optimized for RTX Spark and designed to run alongside your apps. https://t.co/SOiBgeMMjD https://t.co/yPsBt9N58A

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.

Ant open-sources Ling-3.1-flash for sustained work across tools and documents

Ant has open-sourced Ling-3.1-flash, a model with approximately 560B total parameters, about 25B activated per token, and a context window of up to 1M. QbitAI tested it on an e-commerce promotion plan and a simulated medical appointment preparation task, reporting traceable calculations and explicit handling of missing information. The review also found occasional attention drift with long inputs and noted that the current version lacks native visual understanding.

Meta releases Muse Spark Open for coding and agentic work

Meta announced Muse Spark Open, describing it as the most capable US open-weight model to date, built for coding, agentic work, and multi-step reasoning. NVIDIA AI says the 816B model was trained on NVIDIA GB300 NVL72 and has a 1M-token context window for working across large codebases. Meta also produced NVFP4 checkpoints optimized for Blackwell Ultra. https://t.co/yAtoxNeYJm https://t.co/HJUKdC1n3D

The same story is shown in the picks as 《Meta releases Muse Spark Open, an open-weight multimodal mixture-of-experts model》

Meta and Apple team up to run Muse Spark Open on Macs

Meta announced Muse Spark Open today, releasing its weights and saying it can run on two Apple M5 Ultra Mac Studios using the latest MLX framework. Meta says the natively multimodal model has 816 billion total parameters and excels at long-horizon tasks, complex software engineering, and multi-step reasoning. Meta and Apple collaborated over several months, and Apple published a technical post covering MLX, Muse Spark Open, and two other open models.

The same story is shown in the picks as 《Meta releases Muse Spark Open, an open-weight multimodal mixture-of-experts model》

Meta releases Muse Spark Open with downloadable weights and conditional commercial use

Baoyu reports that Meta released Muse Spark Open on October 15, an open-weight model with 816B total parameters, 42B active parameters, and a 1M token context window. It accepts text, images, video, and audio and outputs text; the Muse Model License permits free commercial use subject to attribution, usage restrictions, and a separate agreement for qualifying API businesses. The model card lists minimum memory requirements of 700GB for NVFP4 and 2TB for BF16, while Meta describes optimization for two Mac Studio machines with 512GB unified memory each.

The same story is shown in the picks as 《Meta releases Muse Spark Open, an open-weight multimodal mixture-of-experts model》

Alexandr Wang Says Muse Spark Is the Most Capable US Open-Weight Model to Date

1/ Muse Spark open is the most capable US open-weight model to date. 816B total params, 42B active. 1M-token context. Natively multimodal. Built for coding and agentic work like long-horizon tasks, hard software engineering, and multi-step reasoning. America! 🇺🇲 https://t.co/NjtOB7aY9D

Meta releases Muse Spark Open for coding and agentic work

Today Meta is releasing Muse Spark Open, which it calls the most capable US open-weight model to date. Built for coding and agentic work, Meta says it excels at long-horizon tasks, complex software engineering, and multi-step reasoning. Key model specs include 816B total / 42B active parameters, a 1M token context window, and native multimodality across text, image, video, and audio.

The same story is shown in the picks as 《Meta releases Muse Spark Open, an open-weight multimodal mixture-of-experts model》

MAI announces in-house pre-trained base models in the 3T-parameter weight class

We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI. Post-training results coming soon—stay tuned! Lots of details in our technical blog post. https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/

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.

The Declining Value of CoT Changes How People Work With AI Agents

The sudden collapse of the value of CoT (even summarized CoT) in the agentic era changes working with AI. Effectively steering AIs and getting any sort of observability now requires you to ask for interim outputs & tests, and to use AI help to troubleshoot problems you find.

vLLM Blog Explains Bitwise-Identical Rollout and Training for Faster RL Training

1/ 🚀 New on the vLLM blog: Finish RL Training Faster — Bitwise-Consistent, High-Performance, Agent-Driven Same weights don't guarantee the same policy: inference optimizations can change the logprobs that training recomputes. We make rollout and training bitwise identical https://t.co/S0xkMKfSmc

NVIDIA introduces Nemotron 3.5 Super VL, an open model for long-running AI agents

Meet NVIDIA Nemotron 3.5 Super VL. NVIDIA’s new open model combines reasoning and tool use with text, image, document, and video understanding for long-running AI agents. 120B total parameters with 12B active.

The same story is shown in the picks as 《NVIDIA Expands Nemotron Family With Super VL and Bolt Open Models》

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.

GPT-6-Astra solves 49 of 50 physics research problems at $2.40 apiece

GPT-6-Astra, running in OpenAI's Codex environment without human hints, solved 49 of 50 TheoPhysEval problems across eleven theoretical physics subfields, costing just under $116 overall. Anthropic's Opus 5.5 scored 92 percent, while both models reached 92 percent without internet access. The researchers emphasize that the benchmark tests calculations using known methods, while choosing promising research questions remains a separate challenge for current agents.

StepFun releases open weights for Step 5 Preview, a 600B MoE model

NVIDIA AI congratulates StepFun on Step 5 Preview, calling it StepFun’s largest open weights release. NVIDIA says the 600B MoE model supports agent tasks lasting up to 24 hours and runs at nearly 8K tokens/sec/GPU on NVIDIA GB300 using NVFP4. StepFun’s quoted announcement lists 27B active parameters per token, a 1M-token context window, native text and vision input, and BF16 and FP8 weights on Hugging Face and ModelScope. NVP4 checkpoint: https://t.co/lSECChAwlW Deploy with https://t.co/GO1wqTSBOk https://t.co/28TBzUo6ZU

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.

Shanghai AI Laboratory’s Shusheng-S2 Helps Prove the Three-Dimensional Ghys Conjecture

Xu Disheng’s team used Shanghai AI Laboratory’s Shusheng-S2 model in a human–AI collaboration to prove the three-dimensional Ghys conjecture, with a preprint posted on arXiv. Researchers guided seven rounds of recursive exploration as the problem graph expanded from 15 to 389 nodes. A Peking University team used its SHEAF framework to formalize all four main theorems, producing approximately 550,000 lines of Lean code in about five days and passing Lean checks.