2026-11-20Tencent published the open weights for Hy4, the natively multimodal flagship mixture-of-experts model it launched by API on Nov 4: 770B total parameters with 49B active per token (plus a 1.23B vision encoder and a 3.3B DFly2 speculative-decoding drafter), 78 layers, a 1M-token context window, text, image and video understanding, and computer use. The release includes Hy4 (instruct), Hy4-FP8 and the pre-trained Hy4 base model on Hugging Face, ModelScope and GitCode, with GGUF builds from AngelSlim, day-0 vLLM support and a System Card with safety evaluations. It uses the Tencent Hy License 1.0, which is derived from Apache 2.0 but is not Apache 2.0: model-as-a-service businesses with more than US$50 million in annual revenue or products with more than 100 million monthly active users need a separate written license, while purely internal use is exempt.
Source ↗2026-11-19Microsoft made Aion Instruct (Aion-1.0-Instruct-2.3B), a 2.3B-parameter on-device small language model from its Windows Applied Sciences group, generally available as open weights on Hugging Face under the MIT license and as Microsoft Edge's built-in language model, replacing Phi-4-mini. It handles summarization, rewriting, chat and structured output in eight languages with a 32K-token context. Edge also opened an origin trial of the Prompt API so live sites can use the model on-device, plus a developer preview of an experimental Decisions API. Microsoft says Aion Instruct is also coming to Windows.
Source ↗2026-11-19Ai2 launched Theorizer in Asta, its platform of AI agents for scientific research. Theorizer searches the literature, extracts findings from papers and combines them into testable theories, linking each claim to the supporting studies. Ai2 also open-sourced the code (asta-theorizer, Apache 2.0) so researchers can run it locally with their own model API keys; in Asta, Ai2 covers token costs up to a limit. Scientists have tested it for nearly a year, and Ai2 cautions that it can miss relevant studies and misjudge originality.
Source ↗2026-11-19OVHai LLM (OVHcloud group) released Olala-v1.0-mini-7A1B, the first model in its Olala family, with open weights under Apache 2.0 on Hugging Face. It is a small reasoning model for chat, instruction following and tool use (6.8B total parameters, about 1B active per token, hybrid Mamba-3 / attention mixture-of-experts, 32K context with a 128K version coming soon), trained on three European public supercomputers on 12.8T tokens, and it can run on a laptop or CPU. Base, SFT, summarise and translate variants were released alongside it, and it is hosted on OVHcloud AI Endpoints and Hugging Face Inference Providers.
Source ↗2026-11-18Tencent released open weights for Hy4 flash, the low-cost everyday model it launched by API on Nov 16: Hy4 flash, Hy4 flash-FP8 and Hy4 flash-Base on Hugging Face and GitHub, with GGUF builds from AngelSlim and day-0 vLLM support. The model card lists 220B total parameters (including 64B Engram parameters) with 12B active per token, native text and image input and a 1M-token context window. It is released under the Tencent Hy License 1.0, which is derived from Apache 2.0 but adds a separate-license requirement for large model-as-a-service businesses.
Source ↗2026-11-12The full weights of Inkling 2 Small are now on Hugging Face under Apache 2.0, after external safety testing. It is a 594B-parameter mixture-of-experts model (17B active) with image input, variable effort and a 1M-token context, launched Nov 5 on Tinker with closed weights. It is integrated with Unsloth, SGLang, vLLM and TokenSpeed, can be post-trained with Miles, SkyRL and Snowflake Cortex Training, and is available by API from NVIDIA, Fireworks, Together and others.
Source ↗2026-11-12Open-weight on-device models for laptops, released with Edge0's unified inference framework and deployment tools. Edge0-145B is a sparse mixture-of-experts model with about 10B parameters active per step. Edge0 reports 29GB peak memory, 80+ tokens/s decode and a 128K–256K context on a compatible laptop, with day-one support for macOS, Windows and Linux (Apple M-series and NVIDIA RTX). The benchmarks are Edge0's own (e.g. 64.6 on DeepSWE 1.1). Weights are on Hugging Face (NVFP4 and q3k builds) under the Edge0 Community Model License 1.0, which is not a standard open-source license. An Edge0 App runs offline without an account.
Source ↗2026-11-12Salesforce's first open-weight speech-augmented language model for non-streaming speech recognition, with 26B total and 4B active parameters. Weights are on Hugging Face under Apache 2.0. It targets best-in-class English transcription with multilingual support, entity understanding (phone numbers, emails), robustness to accents and noise, style steering and context biasing. Salesforce says it ranks #1 among open-weight models for non-streaming speech-to-text on Artificial Analysis (#7 overall) as of Nov 12, 2026.
Source ↗2026-11-12pplx-decider-v1.3 and pplx-decider-flash-v1.3 add locate, which extracts verbatim text spans by type or by instructions, alongside yes/no, multiple-choice and rubric decisions. Available in Perplexity's Decisions API, as Apache 2.0 weights on Hugging Face and through OpenRouter.
Perplexity on SIloadout → · Source ↗2026-11-10Open-weight image model built on MiniMax H3 for text-to-image, reference-based generation (up to 16 reference images) and editing, generating native 4K (4096x4096) images in under 10 seconds. Weights are on Hugging Face (MiniMaxAI/MiniMax-H3-Image) under a community license dated Nov 10 that excludes the EU, UK, South Korea and the US; it is also available through the MiniMax API, MiniMax Code and Hailuo AI, with 50% off 4K-and-above requests during launch week.
Source ↗2026-11-09One week after launching LongCat-2.5 (Nov 2), Meituan released its weights on Hugging Face under the MIT License, which allows commercial use, modification and redistribution. The natively multimodal mixture-of-experts model has 1.6T total parameters (about 48B active per token) and targets long-horizon agentic coding, visual coding and 3D tasks. A W4A4 quantized checkpoint is also published.
Source ↗2026-11-09Zyphra released ZONOS3-Mini, a compact zero-shot voice-cloning text-to-speech model: a sparse mixture-of-experts with 100M active and 800M total parameters under Apache 2.0, with controls for identity, accent, mood and word-level delivery. It runs in real time on-device (under 150 ms to first audio on an M3 MacBook Air) with runtimes for browsers via WebGPU, SGLang/vLLM and GGML. Weights are on Hugging Face, and it is also served on Zyphra Inference.
Source ↗2026-11-06Nous Research released Arete-27B, a supervised fine-tune of Alibaba's Qwen3.8-27B (Inkling as teacher) that targets state-aligned framing on politically sensitive topics. On Nous's own 803-prompt eval, answers with propaganda framing fell from 22% to 0.9% with no reported capability regression. Weights are on Hugging Face under Apache 2.0.
Source ↗2026-11-06Video Rebirth released HyperFlow H3 Flash, a ~14B compressed, few-step version of MiniMax-H3 that generates 5–15 second video with synchronized stereo audio in 4–8 steps. INT8 and GGUF builds (Q4_K_M is 9.1 GB) run in ComfyUI on 24 GB+ GPUs. It ships under the MiniMax H3 Community License, which restricts territory (EU, UK, South Korea and US need MiniMax authorization).
Source ↗2026-11-06Kakao released Kanana-2-v-decide-3b-2611, a 3.67B multimodal decision model that returns probabilities over choice, yes/no or score options instead of generating text. It accepts text, JSON or chat with up to 10 images (20,480-token context) and is built on Kanana-2-v-3b. Weights are on Hugging Face under the Kanana Open License.
Source ↗2026-11-05Preferred Networks released PLaMo 3 Translate weights on Hugging Face in four sizes (0.8B, 2B, 8B and 31B), with BF16, quantized, GGUF and MLX builds for on-device translation across 53 languages, plus context and glossary support. Licensed under the PLaMo Community License (free commercial use for organizations with revenue of JPY 1 billion or less, after registration).
Source ↗2026-11-05Baidu released ERNIE-5.5-Flash-Base, the end-of-pretraining checkpoint of a 280B sparse MoE (14B active) trained on 33.5T tokens with up to 128K training context and 1M-token fine-tuning/inference support. Weights (FP8 and BF16) are on Hugging Face under the ERNIE 5.5 Flash Base Model License; there is no mid- or post-training applied.
Source ↗2026-11-04Ricursive Intelligence released Ripple-65B Preview, a 65B dense Transformer pretrained from scratch on 12T tokens with up to 1M context, aimed at agentic coding. Ricursive reports U.S. open-weight SOTA on Terminal-Bench 2.1 (80.52) ahead of much larger open models, trained for about $32M full R&D cost. Weights are on Hugging Face under Apache 2.0 (BF16 instruct + base, NVFP4 variant), with API access on Together AI, GMI Cloud and OpenRouter.
Source ↗2026-11-04Moonshot AI released the full open weights of Kimi K3.1 Pro on Hugging Face as promised: a 2.8T-parameter multimodal MoE (104B active) with a 1M-token context, MoonViT-V2 vision, and MXFP4 weights / MXFP8 activations (~1.56 TB). It is natively agentic and multimodal (text + image). vLLM shipped day-0 support the same day (NVIDIA, AMD, TPU, Alibaba Zhenwu). License is the Kimi K3.1 Pro License (not MIT; large commercial/MaaS use has extra terms). API access continues on platform.kimi.ai; this entry is the weight drop (the Oct 23 API launch is already logged).
Source ↗2026-11-03Sarvam AI released Prism, an open-weight decision model built as a LoRA fine-tune of Sarvam 105B. Instead of generating text, it takes a context, a question and a fixed set of options (yes/no, pick-one or a score) and returns a calibrated probability for every option in a single forward pass, for routing, triage, moderation, judging agent runs and prompt-injection checks. Sarvam reports a 71.8 average across three decision benchmarks with a 44 ms median response on a B200, and support for English, 10 Indian languages and Hinglish. Weights are on Hugging Face under Apache 2.0 (one FP8 GPU with about 120 GB of memory), and it is also available as a hosted Decisions API.
Source ↗2026-11-02Deep Principle, a Hangzhou AI-for-science company, released Hydrogen-1.0, a 753B-parameter open-weight model for chemistry, materials science and life sciences, built by post-training Z.ai's GLM-5.3 on domain data, its own laboratory records and agent discovery trajectories. It is meant to help researchers reason through scientific questions, write and run simulation code, plan experiments and interpret results such as NMR spectra. Deep Principle reports 38.3% on FrontierScience-Research (5 points above GPT-6 Astra) and 93.0% on GPQA Diamond in its own evaluations. Weights are on Hugging Face and ModelScope under the MIT license (BF16, plus NVFP4 and MXFP4 builds), with a 1M-token context and up to 128K output tokens; the model is also available by API through NeuroSpark and inside Deep Principle's Mira AI Scientist.
Source ↗