Thinking Machines, the startup led by former OpenAI CTO Mira Murati, is releasing Inkling 2 Small, a mixture-of-experts model with 594 billion parameters. The company claims it is the strongest western open-weight model, with improved agentic performance. However, this performance comes with significantly higher token usage.
Thinking Machines released Inkling 2 Small, a 594B-total/17B-active multimodal MoE, trained from scratch in 68 hours on 8.4T tokens (NVFP4). It scores 31 on the Artificial Analysis Intelligence Index, beating many same-size Chinese models and ranking high among open-weight models. The model is available on Tinker with Apache 2.0 weights promised next week after safety testing.
Why it matters: This is a major open-weight model release with unprecedented training efficiency and competitive performance, offering concrete technical details and independent benchmarks.
The roundup covers OpenAI's autonomous security pipeline, which it says validated and generated patches for 948 Linux kernel local privilege-escalation vulnerabilities at a $1.2M API cost. It also reports Moonshot releasing Kimi K3.1 Pro open weights, Cognition launching SWE-2 Flash, and Tencent detailing Hy4. Talent moves include Cat Wu leaving Anthropic for OpenAI and Sharon Zhou leaving AMD for Anthropic, while GPT-6.2 Sol and Astra remain unconfirmed rumors.
Why it matters: Readers get a dense, source-linked scan of the week's most consequential model releases, safety findings, infrastructure shifts, and industry moves.
Ricursive Intelligence has released Ripple-65B Preview, a 65B dense open-weight model pretrained from scratch for agentic coding. It achieves 80.5 on Terminal-Bench 2.1, the top self-reported score among US open-weight models, and is available under Apache 2.0 with three checkpoints. The model uses a dense Transformer architecture, trained on 12T tokens with up to 1M context, and claims a full R&D cost of $32M. It trails larger MoE rivals on some benchmarks like SciCode and DeepSWE.
Why it matters: This is a concrete new model release with notable performance claims and open-weight availability, offering practical value for developers and researchers in agentic coding.
Nace AI released Drex Max v1, a dense 27B decision model, as open weights under the Nace.AI Open RAIL-M license. It returns option probabilities in a single forward pass for choice, yes/no and ordinal questions, targeting routing, classification and policy checks. The release includes bf16 weights, a Q8_0 GGUF, an affine Q4 checkpoint and forks of llama.cpp and Ollama, with the 4-bit build fitting 24 GB+ Apple Silicon Macs. Nace self-reports 67.52 on Decision Index 0.4, ahead of Drex v2.0 at 61.25.
Mistral has released the weights for Mistral Large 4 (ML4), a trillion-parameter open-source AI model. Businesses and governments can download and deploy the system on their own infrastructure. The release is positioned toward cybersecurity use cases.
Mistral AI has published the weights of Mistral Large 4, nicknamed Le Chonk, as an open-weight release. The model is described as a 1T parameter multimodal model that Mistral calls the best open-weight model for coding, trained in Europe. The announcement focuses on coding use and defensive or security-oriented applications.
NVIDIA has released Nemotron 3.5 Transcribe, an open-weights speech LLM that handles 40 languages and recordings up to one hour from a single model. The hybrid Mamba-Transformer MoE has roughly 34B total and 4.3B active parameters, with built-in speaker attribution, word timestamps and prompt-selected output styles. NVIDIA reports 4.71% mean WER on nine English Open ASR sets, the best open-weights score, though commercial APIs still lead. Weights ship in BF16, FP8, NVFP4 and GGUF under a commercial-use license, with NVFP4 fitting an RTX 5090.
Why it matters: Developers get a concrete open, commercially licensed speech model that combines transcription, diarization, timestamps and translation, with deployment details and benchmark context.
Hugging Face Bio has released Carbon-X, an Apache 2.0 decoder-only model with 14.8B parameters that processes raw DNA and English in a single token stream and answers in natural language. Built on Qwen3-14B with a hybrid tokenizer adding 5,515 DNA tokens over 6-mers, it was trained in four stages and runs on one 80GB GPU or quantized GGUF builds. Hugging Face reports a 66.7% average across five DNA benchmarks, leading on BioReason Coding at 74.9% but weaker on GenomeQA splice-site questions at 2
Why it matters: It gives readers a concrete, openly licensed DNA model with weight-level access, deployment guidance and self-reported benchmarks for judging practical use.
Moonshot launched Kimi K3.1 Pro on its API, priced at $2.00 per million input tokens and $10 per million output tokens, with a higher-priced "Highspeed" variant that runs 2-3x faster. The model is not yet open weights, but Moonshot promises full weights by Nov 4, 2026. Moonshot's own benchmarks place it near Claude Opus 5.5 and GPT-6 Astra, and Artificial Analysis scored it 55, behind only Opus 5.5 at 58. Simon Willison tested it through the Moonshot API, including a pelican SVG comparison acros
Why it matters: It signals a soon-to-be open-weight model reaching near-frontier benchmark standing at competitive API pricing, which matters to teams choosing models.
IBM released Granite Speech 5.0 2B, a 2B-parameter open speech-language model for transcription and translation in 6 languages under Apache 2.0. It combines a 16-block Conformer speech encoder with a window q-former and a fine-tuned granite-4.0-1b-base backbone, and adds language-ID prompting, keyword biasing up to 2,000 terms, and requested punctuation and capitalization. IBM reports 5.68% average English WER on the Open ASR Leaderboard, ranking second among open-weight models behind Qwen3-ASR-
Why it matters: Developers get a concrete, benchmarked open-weight ASR and translation release with deployment options and licensing details they can evaluate directly.
NVIDIA outlined how European retailers and brands are deploying its open models, inference software and digital twin tooling across commerce and logistics. Shopify post-trained NVIDIA Nemotron 3.5 Lightning for a Sidekick subagent, reporting better accuracy on an internal benchmark alongside 90% lower serving costs and 7x faster responses. Otto Group extended an NVIDIA Omniverse-based warehouse simulation effort to its largest fulfillment center in Germany, and L'Oréal is using NVIDIA ALCHEMI an
NVIDIA announced a broad expansion of its open model family, including Nemotron 3.5 Super VL and Bolt language models, Nemotron Retriever embedding and reranking models, and the new Kumo foundation models for structured business data. The lineup also covers Cosmos 3.5 world models, Alpamayo 2 Edge for autonomous vehicles, BioNeMo RFD4-Proteina for protein design and Ising for quantum computing. The company cites European partners including Mistral, Black Forest Labs and H-Company, plus enterpris
Kyutai, with Inria and École des Ponts, has released REKI, a 2.4B-parameter vision-language-action model for robot manipulation that was pretrained only on public data. It pairs a Qwen3.5-2B backbone with a 200M-parameter flow-matching action expert, and adds memory through 32 learned register tokens cached over the last three frames rather than feeding past frames into the VLM. The authors report 90.4% on LIBERO-Plus and 65.8% on RoboTwin Clean2Rand, with plain-language controls cutting arm osc
Why it matters: It offers a concrete, openly released VLA and a cheap memory design that robotics practitioners can inspect, reproduce and finetune.
DeepSeek has released V4.1-Flash-1017, an open-weight model that reports better benchmarks, faster task completion and shorter outputs than its predecessor. The company also opened its internal search engine through new Search and Fetch APIs priced at $0.75 per 1,000 calls, with free access for now for API users in tools such as Claude Code and Codex. The release targets agentic tasks at lower per-task cost.
MarkTechPost compares 20 decision models released within about four weeks of TypeSafe's Jev launch, spanning Cloudflare, OpenAI, Perplexity, Liquid AI, AWS's Strands Labs and independent developers. The guide lists size, license, input modalities, Jev API compatibility, pricing and reported benchmarks for each. Most figures are vendor-reported and per the article are not comparable across benchmark suites, while one third-party run from Respan keeps Jev ahead on accuracy but shows Tev1 4B better
China plans to restrict exports of 23 key AI technologies, including model weights, source code, and training methods for systems meeting thresholds of 10^26 FLOPs or 2 trillion parameters. A fast-track rule would still clear open-source releases once registered. The report suggests the move may push Western companies toward US open-weight models over Chinese alternatives.
Reflection AI has released Beam, an open-weight sparse mixture-of-experts model built for agentic coding and long-horizon software tasks. Beam has 501 billion total parameters with 23 billion active, a reported 1 million-token effective context length, and Apache 2.0 weights on Hugging Face in BF16, FP8 and NVFP4 checkpoints. The AWS post details deployment via SageMaker JumpStart real-time endpoints or SageMaker HyperPod Inference with JumpStart, SGLang or vLLM, including GPU sizing and tuning
Why it matters: It documents a new open-weight agentic coding model with concrete parameter counts, licensing and deployable AWS serving paths that developers can evaluate.
