openmodelweights.com

Today · Fri 20 Nov

“Collaboration over competition”: Hugging Face hands model inference traffic to OpenRouter

AI-generated summary Hugging Face and OpenRouter announced a partnership to route language-model requests using OpenRouter infrastructure. The integration is live and covers 160 models from 38 AI labs.

Also covered byaipractice.osuslova.proWritebleOpen Model Weights
Earlier:

Hugging Face Qwen3.8 Pi LoRA Improves Coding-Agent Efficiency

Original minimalist editorial illustration representing code in AI research; conceptual, not actual model output.

AI-generated summary Open Model Weights reports the release of an Apache-2.0-declared LoRA adapter for Qwen3.8-27B trained inside the pi coding-agent environment. Publisher evaluation data indicates that SWE-bench Pro scores rose from 39.2 percent to 64.7 percent on selected hard tasks.

Also covered bye/accaiagentnews.io
Earlier:

Tuesday 17 November

Nemotron for Science: Open Chemistry Data and Scientific Agents

Original minimalist editorial illustration representing science in AI research; conceptual, not actual model output.

AI-generated summary NVIDIA launched Nemotron for Science to help research teams adapt open artificial intelligence models to chemistry, biology and laboratory workflows. The release includes LitmusBench v0.2, a chemistry resource with nearly 50,000 questions drawing on molecular-property prediction and drug-discovery data.

Also covered bySoBasically

LongCat-Video-Avatar 1.6: MIT Weights and Audio-Driven Video

Original minimalist editorial illustration representing audio in AI research; conceptual, not actual model output.

AI-generated summary Meituan released LongCat-Video-Avatar 1.6 on November 15 to make audio-driven synthetic video more natural and stable across extended sequences. Meituan says its continuation pipeline supports video exceeding ten minutes with reduced error accumulation.

Also covered byGitSeonjeap
Earlier:

Tess 4.5: Open 327 MB Coding LoRA for Inkling 2 Small

Original midnight violet editorial illustration of a glowing amber adapter prism connected to a large glass neural matrix with fine code-like light trails.

AI-generated summary Continuity Research developer Migel Tissera has released Tess-4.5-594B-A17B, which is a coding-focused derivative of Thinking Machines' Inkling 2 Small. The developer reports training on 46.9 million tokens from anonymized engineering sessions and releasing a separate LoRA adapter of about 327 MB alongside a full merged checkpoint.

Earlier:

Huawei Cloud open-sources its CloudRobo-WM1.2 robot world model, and the price of entry is Ascend compute

Huawei Cloud open-sources its CloudRobo-WM1.2 robot world model, and the price of entry is Ascend compute

AI-generated summary Huawei Cloud released the CloudRobo-WM1.2 embodied world model, which was trained on Ascend chips and features a DualAct architecture. Huawei says it will open-source the model weights, inference stack, and code while leaving robot manufacturing to its partners.

Also covered byOpen Model Weights
Earlier:

Monday 16 November

NVIDIA Cosmos 3: Four Open-Weight World Models

Original conceptual artwork of a luminous autonomous vehicle path, layered camera panels and cyan LiDAR scanning rings inside a three-dimensional simulated environment.

AI-generated summary NVIDIA published four new Cosmos 3 Nano open-weight checkpoints featuring sixteen billion trainable parameters in a technical release. The systems include Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer, and Cosmos3-Nano-Sim-Bimanual under OpenMDW 1.1.

Earlier:

Arete-27B: Open-Weight Alignment Fine-Tuning Examined

Original atmospheric visualization of a 27-billion-parameter model represented by two diverging luminous neural pathways and a central evidence lens on a violet graphite background.

AI-generated summary Nous Research has released Arete-27B, a supervised fine-tune of Alibaba's Qwen3.8-27B designed to examine how an additional training phase alters responses to politically sensitive questions. Nous says its internal rubric raised mean factual recall from 2.32 to 2.98 while the proportion of responses adopting Chinese official framing fell from 22 percent to roughly 0.9 percent.

Transformers 5.23: Inkling 2, LFM3 and Model Exports

Original cinematic technology illustration showing a glowing modular transformer network with distinct export paths to three floating compute chips in navy, teal and cyan.

AI-generated summary Hugging Face released Transformers 5.23.0 on November 13 with added implementations for Thinking Machines Lab's Inkling 2, Liquid AI's LFM3, and other open-weight architectures. The software update also introduces the ExportedModel path to unify backend-specific model export workflows for tools including ONNX Runtime, OpenVINO, and ExecuTorch.

Baseten open-sources UltraFast engine for DeepSeek-V4.1-Flash

Baseten open-sources UltraFast engine for DeepSeek-V4.1-Flash

AI-generated summary Baseten has open-sourced the UltraFast inference engine optimized for NVIDIA B200 and B300 GPUs to support DeepSeek-V4.1-Flash. The Rust and CUDA engine achieves decoding speeds exceeding 2000 tokens per second using speculative decoding.

Also covered byOpen Model Weights
Earlier:

Tencent to Release Open Weights for Hy4 Flash AI Model

Tencent to Release Open Weights for Hy4 Flash AI Model

AI-generated summary Tencent announced plans to release the open weights for its Hy4 flash large language model in the near future. Tencent claims the multimodal model offers productivity performance parity with Claude Haiku 5.5 at a task cost that is 18 times lower.

Also covered byOpen Model Weights
Earlier: