Tencent launches Hy4 flash with 12B active parameters, but open weights remain pending
The new 224B multimodal model promises cheaper agent workflows; Tencent has opened commercial API access, not a verifiable public weight release.
OPEN MODEL WEIGHTS · DAILY INTELLIGENCE
Source-first briefs on model releases, weights, licenses, runtimes, training infrastructure and ecosystem shifts — tied back to the evidence standards of the registry.
Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.
LATEST EDITION · 2026-11-16
The new 224B multimodal model promises cheaper agent workflows; Tencent has opened commercial API access, not a verifiable public weight release.
The Apache-2.0 base checkpoint listens and speaks simultaneously, with openly available local inference code and carefully qualified latency claims.
A public LoRA and merged checkpoint adapt a huge sparse model for software agents; publisher benchmarks suggest a narrow gain, not universal superiority.
Hugging Face expands native open-weight architecture support and introduces an ExportedModel interface for moving compatible checkpoints into supported inference runtimes.
Cosmos3-Nano variants generate multi-view driving observations, simulator-conditioned video and robotic-action forecasts; licenses and physical fidelity still require scrutiny.
A Qwen3.8-27B supervised fine-tune exposes how post-training changes political response framing, but the reported improvements depend on the authors' own rubric and judge.
Baseten releases a Rust/CUDA serving engine for DeepSeek-V4.1-Flash. Its Blackwell throughput is vendor-reported, and Elastic License 2.0 restricts managed hosting.
Updated quantized weight files target better local model accuracy at similar sizes; Unsloth's 5–10% comparison remains a publisher-reported benchmark claim.
Google outlines local ASR and TTS stacks around Parakeet, PocketTTS, Phonon and Kokoro, with separate Android-system and cross-platform model paths.
The local-inference stack gives downstream C/C++ integrations an explicit major, minor and patch compatibility policy; it is not a new model checkpoint.
The openly documented 8,115-question benchmark tests whether AI helpers recognize a mistaken proposed solution and redirect toward the user's actual goal.
Researchers add independently trained experts to a frozen Qwen3.5 MoE and combine the skills in one checkpoint, avoiding a standard adapter-merging bottleneck.
Four public GGUF/NVFP4 repositories target NVIDIA GPUs, but host-memory offload and custom model licenses make the practical deployment story more nuanced.
A reproducible ModernBERT example trains token embeddings to survive pooling, improving NanoBEIR retrieval quality at the same document-vector budget.
A new public benchmark evaluates 21 decision and reasoning systems across 11 domains and separately tests option-order stability and missing-answer recognition.
Perplexity describes two Apache-licensed decision checkpoints for classification and routing, but the advertised Hugging Face weight access needs a separate check.
An open training recipe adds local and hosted fine-tuning paths for TRELLIS.2 geometry, while gated DINOv3 access and substantial GPU needs still matter.
A newly merged fix restricts server-side request paths involving the browser CORS proxy and remote media, highlighting security considerations in local inference.
The November 12 checkpoint release turns a hosted-only multimodal agent model into downloadable Apache-2.0 weights, but local operation still requires a GPU cluster.
The 26.6B-total speech model puts transcription style and critical entity capture in focus; benchmark rankings and throughput remain issuer-reported.
Two Apache-2.0 experimental checkpoints encode instruction-to-data bindings inside the chat template, offering a testable alternative to plain-text tool context.
The compact sparse speech model ships with Apache-2.0 weights, multiple local runtimes and a new evaluation suite; latency and quality rankings are Zyphra's own measurements.
The dense model pairs 18 recurrent-attention layers with six full-attention layers; its 1.5B effective-parameter description is not its total checkpoint size.
The GLM-derived model targets longer tool-driven workflows; first-party benchmark scores are model-level claims, not independent tests of the published FP8 export.
A publicly downloadable video-action policy now comes with base, DROID and YAM checkpoints, but its robotics success rates remain publisher-reported.
QVAC's Qwen-derived model prioritizes on-device tool execution, but its benchmark comparisons and prompt-injection safety figures have not been independently reproduced.
Simons supplies four probabilistic forecasting checkpoints and unusually broad research materials, but its weight, code and dataset licenses differ.
Public MIT-licensed artifacts bring Meituan's enormous vision-and-code mixture of experts into the open, but reproducible deployment and evaluation remain substantial hurdles.
PAN 2 combines a language-based event planner with a 29B video diffusion backbone to explore possible action outcomes, while physical fidelity still needs controlled tests.
The new image checkpoint unifies creation and multi-image editing at high resolution, but the accompanying text encoder, hardware footprint and custom license are decisive.
The Polish-centered open-weight model combines European multilingual coverage with instruction, reasoning and structured tool-call support, but file gating and real-language testing matter.
A GLM-5.3-based scientific model offers downloadable weights and hardware-specific variants, while its laboratory claims remain publisher-reported.
Gradium's technical deep dive clarifies how tiny offline speech models work and why the commercially usable 31M release differs from the 40M family.
A new evaluation asks whether language models can follow evolving entities and assignments, rather than simply retrieving a fact from a long input.
The new serving release spans oversized mixture-of-experts models, fused low-latency kernels and wider accelerator support.
New serving paths span large multimodal checkpoints and speculative decoders, while a CUDA compatibility change matters to production upgrades.
An API-compatibility fix changes a subtle but consequential source of output variability for locally served open-weight models.
A 35B-total, 3B-active model adds block diffusion to open-weight coding agents, with substantial publisher-only speed claims.
Two source-published models score typed options in one pass instead of generating sentences, but calibration claims require independent testing.
A compact document-specialist checkpoint brings tables, formulas and structured extraction to local open-weight systems.
A downloadable four-billion-parameter vision-language policy combines screenshot-grounded clicks with persistent terminal operations instead of routing through separate agents.
The new quantized checkpoint changes expert arithmetic and compresses Engram tables while imposing an important patched-vLLM requirement.
Datalab reports better tables, multilingual layout reading and fewer decoding loops, but the model weights are not unrestricted Apache-2.0 assets.
A downloadable Unitree G1 policy turns changing motion goals into whole-body actions, with distinct native and browser artifacts and clear deployment limits.
The new framework release adds native pathways for two newly published models while introducing FP8 backpropagation and a more economical MoE adapter-training path.
Liquid AI's local model family reaches Linux and Windows in Ollama, including speech-focused and multimodal variants previously introduced on Apple hardware.
A newly downloadable trillion-parameter multimodal checkpoint earns a Leading listing, with its source revision, server-scale demands and license conditions clearly separated.
NVIDIA's audio-specialist checkpoint earns a Notable listing for offline multilingual transcripts, translation and speaker-attributed word timing.
Hy-SimulMT brings downloadable Chinese-English live-translation checkpoints, with real-time spoken output depending on a separately configured TTS system.
The new d3-ultra weights return typed choices, scores, visual boxes and tracks rather than natural-language completions.
A high-severity security patch closes a model-configuration import path that could execute repository code even with remote-code trust disabled.
New ggml-org changes add model conversion and runtime pathways for multiple LFM3 modalities without introducing new model checkpoints.
Liquid AI opens a six-model LFM3 family built around a shared 1.7B backbone, with specialized vision, audio, omni and decision variants.
Unlimited OCR 2 publishes MIT-listed vision-language weights and a sparse-reference approach intended to keep long-document parsing manageable.
FrontiersMind publishes Apache-2.0 tools to turn supported model backbones into constrained decision engines with measurable output probabilities.
Nanbeige publishes an Apache-2.0 dense 33B checkpoint for coding and agent workflows, positioning it as the teacher for a smaller future model.
A new technical report separates streaming transcription, speaker-labelled text and audio understanding into distinct Qwen-Audio 3.1 deployment profiles.
A new SGLang design lets hybrid attention models share GPU cache capacity, with a carefully bounded publisher test showing improved prefix retention.
The ggml project merged a DeepSeek V4.1 Flash model-support change, extending local-inference code while leaving hardware and quality validation to users.
A Korean gaming company publishes a hybrid-attention sparse model with public preview weights, long context and separately documented inference variants.
A new 40B-class sparse model pairs downloadable Apache-2.0 weights with a dedicated agentic-coding and inference story; speed comparisons remain vendor claims.
The local model runner now offers macOS speech generation and d3 classification or scoring from prompts, structured state and images.
A Python client release improves data-pipeline freshness and job visibility, while preserving a clear boundary between Bucket updates and model-repository revisions.
New Hi3D checkpoints refine coarse meshes from reference images; the 1536³ label is a sparse grid target, not dense voxel memory.
OlmoEarth v1.3 offers open model weights and precomputed planetary vectors, reducing the need to rerun satellite encoders for every mapping task.
The open training roadmap moves from imitation traces to reward-driven, multi-turn agent environments without claiming a new model-weight release.
Carbon-X places genomic sequences and English in one decoder, but publisher benchmark wins are not laboratory validation.
A new 27B ternary-weight serving bundle verifies drafted tokens against the original model while targeting compact local inference.
Laghu couples a small local chat model with checkpoints, optimizer states and evaluation records instead of presenting size alone as proof of quality.
The new 27B and 4B Apache-licensed checkpoints turn text and images into typed decision probabilities, with different accuracy and serving trade-offs.
A Europe-specific Earth-2 nowcasting workflow uses Meteosat observations to predict cloud evolution, while satellite licensing and forecast skill need scrutiny.
A domain-specific tokenizer can matter as much as model size for repeated structured outputs, but the reported saving is one team's preliminary case study.
IBM's Apache-licensed speech release offers three routes for transcription, translation and speaker-labelled audio, with distinct deployment trade-offs.
NVIDIA's new embed-and-rerank family targets visual documents and agent evidence retrieval; its OpenMDW 1.1 license needs separate review.
Kumo Forecast pairs pretrained temporal representations with historical-neighbor retrieval and separate interpretability outputs.
Action-conditioned surgical video models show a new direction for Physical AI, but clinical safety and simulation fidelity remain unproven.
TorchTPU gives open-model teams a familiar PyTorch path onto Google accelerators while leaving model-specific compatibility and actual cost to be tested.
New MPS kernels and FP8 support bring notable publisher-measured gains for some Apple Silicon open-model inference and video workloads.
The Swift framework reduces switching friction between GGUF, MLX, Core ML and hosted services, while individual model capabilities and licensing still differ.
A 631M-parameter document parser extracts layout, text and bounding boxes from page images, with producer-reported quality tests and OpenMDW license terms.
The TorchNeuron public beta promises less specialized code for training open weights on AWS Trainium, but dynamic-shape and runtime limits still matter.
New sparse embeddings offer English and multilingual semantic matching without requiring all the token-level index storage of late-interaction retrieval.
A small open speech recognizer targets transcription without uploading audio, with different laptop and phone language coverage and device-specific tradeoffs.
Inco AI extends parallel draft decoding to Qwen3.8 Flash Next, with a new auxiliary checkpoint rather than new target-model weights.
Tencent releases a 30B-total, three-billion-active model trained to decide what should enter a persistent agent memory bank.
Four small forecasting checkpoints extend the open-weight ecosystem beyond chat models, with benchmark scores and license terms requiring qualification.
The October 19 serving release advertises a 10x loader improvement and new model integrations; it also fixes a speculative-decoding state error and raises the Python minimum.
The new GoTo-AI checkpoint targets Indonesian culture and regional-language gaps, but its Gemma terms and full-precision memory footprint need careful reading.
StepFun's downloadable multimodal MoE targets long-horizon agents, but its million-token context claim and Apache-style license language require precise qualification.
The compact Bolt checkpoint is designed for RTX and DGX Spark with native low-bit weights, while its BF16 sibling serves a different adaptation purpose.
The new 120B-total, 12B-active vision-language parent of Nemotron Bolt targets document-heavy agents and custom training, with model-specific GPU requirements.
The October 18 serving release adds a structured decision endpoint and broader low-precision caches, while tightening security defaults and breaking some existing deployments.
OpenBMB's Apache-2.0 vision-language family pairs a 35B-total, 3B-active MoE with a compact 1B alternative and an unusual expert-prefetching approach for phones.
The stable Ollama update adds parsing for a new low-bit GGUF type and Windows runner cleanup, with smaller changes to a cloud model and interactive prompting.
The 560B-total, 25B-active hybrid reasoning model pairs KDA and gated MLA attention with sparse experts; the public MIT checkpoint is aimed at server-scale agent and coding workloads.
Two downloadable FLUX.3 Image Klein checkpoints support text-to-image, reference-based editing and bounding-box placement, but their Kommunity License is not a blanket commercial-use grant.
DMIS Lab publishes three Apache-2.0 checkpoints that combine dual masking objectives with a 32K context, aiming to serve classification and retrieval without two separate encoder families.
The October 17 checkpoint is a distinct MIT-licensed release; DeepSeek reports higher agent and coding scores, but reproducibility and prompt encoding matter.
The Apache-2.0 8B and 30B checkpoints package decision and safety adapters with an English speech encoder; IBM's performance claims remain unverified here.
New October technical documentation explains how the previously released 8B and 70B weights handle image and audio input, with notable transparency and access caveats.
A newly merged, opt-in Discover Models interface helps local-model users browse GGUF quantizations, resume partial downloads and manage speculative draft files.
Meta's 816B-total, 42B-active mixture-of-experts checkpoint accepts text, images, audio and video, but its gated weights and custom license require careful deployment review.
Beam moves from an early-October announcement to downloadable 501B-total, 23B-active model weights in three numerical formats, plus an optional speculative-decoding drafter.
The experimental 4B-parameter Dory model reuses a recurrent transformer block at inference time, offering adjustable loop counts rather than a larger checkpoint for each compute budget.
Mistral's first time-series model family spans 5M, 28M, 123M and 209M parameters, supports covariates and probabilistic outputs, and is downloadable under modified MIT terms.
The new 24B-total, 2B-active mixture-of-experts checkpoint offers 131K context and tool-use training; Liquid AI's speed and benchmark claims remain publisher-attributed.
Ai2 publishes HiRO-ACEv2 weights and inference guidance for global climate emulation and 3 km downscaling; its reported cost and extreme-event fidelity remain author measurements.
AI-generated editorial illustrationThe University of Innsbruck team's 179M-parameter English retriever packages dense and token-level matching into one ColBERT-compatible model; its BEIR numbers are author-reported.
AI-generated editorial illustrationSparseup-fr-v0 adapts Linkup's sparse-retrieval recipe to French with a 50,240-dimensional output and a downloadable model card; benchmark gains are author-reported.
AI-generated editorial illustrationThe vLLM maintainers outline Q4 work on faster agentic serving, production recipes, model support and reliability; the roadmap is not a release announcement.
AI-generated editorial illustrationLightOn’s local 2B code-search agent returns source locations to coding assistants; the publisher reports fewer paid tokens, with latency and accuracy trade-offs.
AI-generated editorial illustrationThe 397B-total, 17B-active Hunmin-CUA checkpoint targets desktop and browser agents using low-rank transfer from Qwen-CUA followed by post-training.
AI-generated editorial illustrationLightOnOCR-3 adds labeled document regions, image descriptions and chart extraction to lightweight OCR, with downloadable 0.8B, 1B and 4B weights.
Tencent's 5B multimodal parser emits structured JSON across documents, images, audio and video, but its custom weight license expressly says it is not intended for EU use.
Ai2 describes replacing priority queues with GPU-time budgets, fair-share scheduling and protected run windows; its operational results are reported by the lab itself.
Perplexity publishes MIT-licensed 0.6B and 9B late-interaction embeddings for text, images and documents, with a shared index-query space and explicit storage tradeoffs.
Cloudflare's new Apache-2.0 Clef-omni checkpoint scores structured decisions across audio, video, images and text; publisher latency claims are not independent OMW tests.
Qwen releases a downloadable 7B visual-generation checkpoint with a configured eight-step schedule, while its research license restricts commercial self-hosting.
FINAL-Bench releases a Safetensors classifier that produces typed decision probabilities without generative decoding; its creators have corrected initially inconsistent per-type benchmark figures.
JetBrains publishes an Apache-2.0 MoE checkpoint with 2.5B active parameters, 131K context and reinforcement-learning upgrades for repository-level agent tasks; performance claims remain publisher-attributed.
CIQ's orchestration release lists eleven validated model presets, a vLLM catalog path and autoscaled endpoints across private and cloud GPU fleets; it does not announce new model weights.
Liquid AI’s d1-3B and experimental d1-omni-600M models score typed decisions instead of generating prose; its launch includes published model cards, architecture details and hardware-specific latency claims.
An NVIDIA research report describes task-specific Nemotron checkpoints, curated training sets and generate–verify–refine pipelines for competitive programming and mathematics, with performance claims tied to their exact evaluation conditions.
The 3.35B open-weight research model is trained to keep reasoning in the user’s language, combining English reasoning, translated multilingual reasoning and non-reasoning multilingual data.
The 740M-parameter Apache-2.0 model maps text, code, images, video and audio into one embedding space, with modular encoders, an 8K context window and published on-device deployment paths.
Hugging Face’s October 6 release adds Google’s multimodal EmbeddingGemma 2 architecture, makes token-dispatch expert parallelism the default for Qwen3 MoE and Mellum, and repairs quantized-cache behavior.
The October 5 serving release adds a GPU-resident weight-cache daemon, experimental initialized-engine snapshots, new speculative-decoding paths and a broad set of DeepSeek V4.1 Flash optimizations.
The official Index-Translate GGUF repository packages llama.cpp conversions from Q2_K through F16 for the 36B-parameter preview model, with Apache-2.0 metadata and local-serving instructions.
CoreAI’s pipeline turns observed agent failures into executable training tasks, uses a stronger teacher to identify solvable gaps, and validates prompts, environment state, trajectories and verifiers before samples enter training.
The preliminary RIDM release includes separate optical-flow and pixel models that infer camera motion from video, with Safetensors checkpoints, reproducible inference code and explicit Apache-2.0 weight licensing.
NVIDIA’s latest telecom report says 89% of respondents view open-source models and software as important to AI strategy, while a new 30B Nemotron telco model and an open fine-tuning recipe target operator workloads.
The 27B Clef and 9B Clef-Flash turn text, JSON, images or video into typed probability outputs, with Apache-2.0 repositories, sharded Safetensors and Jev/SystemOne-compatible interfaces.
The quantized decision-model checkpoint cuts publisher-reported weight download size from 49.5 GB to 18 GB and weight memory to 17.1 GiB while retaining a Gemma 4 backbone and a separate System 1 decision layer.
The multimodal MoE activates 49B parameters, lists a 1M-token context window and is available first as an API preview; Mistral says downloadable weights will arrive by month-end.
The new Falcon family spans a 7B Emirati-dialect language model, a 1.6B speech recognizer and a compact Arabic OCR model, extending TII's stack across text, audio and documents.
Beam is a 501B-total, 23B-active sparse MoE for coding and agentic workloads, but the current release is an early-access preview rather than a downloadable checkpoint.
The runtime release combines a Hugging Face Hub download pipeline with new batch APIs, decision-model serving and explicit support for GLM-5.3-Flash, Clef and Qwen4Exp workflows.
The 78.1B-parameter mixture-of-experts model targets German-English sovereign deployment, with FP8 and BF16 weights, tool calling and a stated context ceiling of 1,048,576 tokens.
AstaBrief turns retrieved scientific evidence into cited reports and gives institutions an open-weight option for running the report-writing stage on their own infrastructure.
The local inference stack can now score typed options directly instead of generating free-form text, opening a different deployment path for routing, moderation and agent control.
Ai2 redesigned its open training infrastructure around MoE routing, expert parallelism and lower-precision execution, with the next generation of Olmo expected to use the stack.
The Kumo Structured family targets classification and regression directly from labeled tables, with three small model sizes and a commercial-use license.
H Company’s new agentic family is designed to move between GUIs, code, MCP and APIs, with downloadable weights in multiple precisions and public benchmark trajectories.
The DSpark companion model adds speculative decoding to Liquid AI’s vision-language stack with integrations for llama.cpp, MLX-VLM and SGLang.
FLUX 3 Action predicts future frames and actions together and arrives with robot-policy checkpoints, LeRobot integration and a dedicated weights license.
Nemotron 3 Diarization targets real-time and offline speaker attribution, turning speech streams into speaker-aware timelines for meetings, calls and voice agents.
GGUF checkpoints can now be loaded through familiar Transformers APIs while reusing ggml kernels to target practical local inference, initially with an Apple Silicon focus.
The $12.93 billion agreement would put the central model-sharing platform inside NVIDIA while the companies publicly commit to model, cloud and accelerator choice.
EDITORIAL STANDARD
Publisher blogs, model cards, release notes and repositories are preferred over second-hand coverage.
Benchmark, speed and performance claims remain publisher claims unless Open Model Weights independently measures them.
The system can publish up to ten briefs per day, but fewer are published when there are fewer meaningful developments.
Each brief distinguishes a model-record candidate from runtime, compatibility, infrastructure or ecosystem news.
Open Weight Intelligence is produced with automated research assistance and source-first editorial rules. Every brief links to its primary source, and unsupported details remain absent rather than inferred.