Skip to content
K = 0.731
2026.10.17 · Sat

#开源生态

#开源生态

Models

New models and versions, open weights, and changes in capability and pricing, with their evaluations.

10/16Fri
  1. Hugging Face Blog83

    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.

    Why it matters: The deployment guidance distinguishes active computation from total weight memory, helping teams assess hardware requirements for running and adapting a sparse model with a long context window.

  2. NVIDIA AI · X77

    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.

    Why it matters: The Apache 2.0 license, multiple weight formats, and released training harness give developers concrete information for assessing Beam’s suitability for their deployment workflows.

  3. Hugging Face Blog87

    Meta releases Muse Spark Open, an open-weight multimodal mixture-of-experts model

    Meta’s Muse Spark Open is an open-weight mixture-of-experts model with approximately 816 billion parameters, around 42 billion active per token, and a context window of up to one million tokens. Trained for agentic coding, it accepts text, images, audio, and video and generates text. Hugging Face provides inference and deployment recipes, reports its own quantization tests, and demonstrates attention-layer LoRA fine-tuning with TRL while keeping the experts frozen.

    Why it matters: The deployment recipes clarify how sparse expert activation reduces per-token computation while retaining substantial memory requirements, helping developers assess hardware needs across checkpoint formats and serving frameworks.

  4. The Decoder78

    Black Forest Labs releases FLUX 3 Image Klein with 8 GB VRAM support

    Black Forest Labs has released FLUX 3 Image Klein, a 5.6-billion-parameter open-weights model for text-to-image generation, image editing, and generation with up to ten reference images. It supports native 4K output, and its INT8 build runs on NVIDIA GPUs with 8 GB of VRAM, generating an image in under 9 seconds on an RTX 4070. The FLUX Kommunity License allows free commercial use for companies making under $5 million annually.

    Why it matters: The hardware requirements and commercial license threshold provide concrete context for assessing whether local image generation fits a team's available GPUs and business size.

  5. The Decoder81

    Nvidia releases Nemotron 3.5 Super VL for continuously running enterprise AI agents

    Nvidia released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters and 12 billion active per token, for continuously running enterprise agents. Nvidia says it approaches Nemotron 3 Ultra’s accuracy on specified agent tasks with about one-fifth the parameters, and is releasing weights, datasets, and training recipes. The accompanying Nemotron 3.5 Bolt 74B-A9B 24GB targets local agents, while Bolt E4B 4GB is planned for later this month.

    Why it matters: The deployment details clarify how quantization, caching, and hardware requirements shape the practical tradeoffs of running continuous enterprise agents on company infrastructure.

10/15Thu
  1. NVIDIA Blog81

    NVIDIA Expands Nemotron Family With Super VL and Bolt Open Models

    NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model for multimodal agents, and Nemotron 3.5 Bolt 74B-A9B 24GB for local agentic workloads; Bolt E4B 4GB is coming Oct. 23. Super VL targets tool use, business workflows and visual document understanding, with open weights, datasets and training recipes. Bolt 74B-A9B uses native 2-bit quantization to fit in 24GB of memory, while Bolt E4B targets physical and conversational agents.

    Why it matters: Open weights, datasets and training recipes give organizations a path to adapting agents to their own workflows while meeting local deployment and data residency requirements.

  2. NVIDIA AI · X76

    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

    Why it matters: The reported throughput ties Step 5 Preview to a specific GPU and precision format, giving readers concrete deployment context for its long running agent workloads.

  3. Hugging Face Blog74

    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.

    Why it matters: Published intermediate checkpoints, training mixtures, and evaluation settings provide a concrete basis for investigating how individual training stages change model behavior and for reproducing reported results.

  4. 雷锋网83

    Kunlun Tech releases and open-sources Matrix-Game 4.0 for single-GPU real-time gameplay

    Kunlun Tech released Matrix-Game 4.0, reporting 480p gameplay at 60 FPS on a single NVIDIA RTX 5090 with end-to-end response latency below 150 milliseconds. Its game engine maintains world state and interaction rules, a coding agent modifies game logic, and a video model generates detailed visuals from structural frames and dynamic semantic conditions. Code, weights, and simplified scene models are available at https://huggingface.co/Skywork/Matrix-Game-4.0.

    Why it matters: The architecture offers a transferable approach to interactive generation by assigning persistent state and rules to an engine while using a video model for visual rendering.

10/14Wed
  1. Artificial Analysis · X78

    Fish Audio’s Drama 3 Preview leads open-weights models in Artificial Analysis TTS rankings

    Artificial Analysis reports that Fish Audio’s Drama 3 Preview ranks #1 among open-weights models on all three of its Text to Speech leaderboards and leads all models in Pronunciation Robustness. It scores 93.2% on pronunciation robustness, ahead of Eleven v4 at 91.7% and Eleven v4 Turbo at 90.1%, while ranking #10 in Provider Voice and #14 in Controlled Voice. Its weights are available on Hugging Face under the Fish Audio Community License, and Fish Audio API pricing is $15/1M characters.

    Why it matters: The comparison separates pronunciation robustness from voice preference rankings, showing how Drama 3 Preview’s lead on one measure differs from its placement on others.

  2. Hugging Face Blog79

    Liquid AI releases LFM2.5-24B-A2B for fast, private on-device agents

    Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning model with 24B total parameters and 2B active parameters, for on-device agents that reason and call tools. Its context window expands from 32K to 131K tokens, and both the reasoning and Base versions are available on Hugging Face. Liquid AI reports 184 tokens/s on an Apple M5 Max and competitive tool-use results, while acknowledging that larger models retain advantages on knowledge-heavy and coding benchmarks.

    Why it matters: The benchmark comparison helps distinguish workloads suited to local deployment, with strong tool use and instruction following alongside weaker knowledge and coding results than some larger models.

10/9Fri
  1. Qwen · X65

    Qwen releases Qwen-Image-2.1-Turbo, generating and editing images with 8-step denoising and open weights

    Qwen has released Qwen-Image-2.1-Turbo, based on the same 7B visual generation architecture as Qwen-Image-2.1, requiring only 8 denoising steps to generate and edit images, with weights now open on ModelScope and Hugging Face.

    Why it matters: Qwen's image model gains a Turbo accelerated version, with 8-step sampling and open weights offering a generation and editing path that can be integrated directly.

10/8Thu
  1. Hugging Face Blog62

    Liquid AI open-sources d1-3B and d1-omni-600M on-device decision models

    Liquid AI released two open-weight models in its d1 decision-model family: d1-3B and the experimental d1-omni-600M.

    Why it matters: Liquid AI open-sourced two decision models and reported on-device latency and results across seven benchmarks, providing evidence on the use of small models for structured decisions.

10/7Wed
  1. Google DeepMind71

    Google DeepMind releases EmbeddingGemma 2, a lightweight open multimodal embedding model

    Google DeepMind released EmbeddingGemma 2, a 740M-parameter open model based on the Gemma 4 architecture and licensed under Apache 2.0. It maps text, images, audio, and video into a shared embedding space.

    Why it matters: EmbeddingGemma 2 maps text, images, audio, and video into one embedding space and reports concrete on-device memory and vector truncation figures, helping assess local multimodal retrieval.

10/6Tue
  1. Mistral AI80

    Mistral launches public preview of Mistral Large 4, a 1-trillion-parameter multimodal model

    Mistral released a public preview of Mistral Large 4, a natively multimodal model with 1 trillion parameters and 49 billion active parameters. Its weights will be opened at the end of the month.

    Why it matters: Mistral outlined the open-weight release schedule for its 1-trillion-parameter multimodal model and its investment in European compute, allowing comparison with mainstream open models.

  2. Ollama · X65

    Reflection AI announces Beam, a 501B-parameter agentic open model

    Reflection AI introduced Beam, described as a highly efficient agentic open model with 501B total parameters and 23B active parameters. The company says it advances the Western open frontier on coding and agentic tasks and was trained end-to-end from scratch, with full weights to be released this month. Ollama said more US models are coming to its platform.

    Why it matters: Reflection AI's Beam is presented as an efficient agentic open model with 501B total and 23B active parameters, giving readers a concrete comparison point for open-weight coding and agentic models.

10/1Thu
  1. Hugging Face Blog62

    AI2 releases Olmo-core 3, open and scalable training infrastructure for large MoE models

    AI2 released Olmo-core 3, a major upgrade to its large language model development framework. It redesigns the open MoE training system to scale MoE training to trillions of parameters while maintaining compute efficiency.

    Why it matters: The source provides throughput and memory comparisons for the MoE training stack, helping readers assess changes in the cost of large-scale MoE training.

9/29Tue
  1. Hugging Face Blog62

    NVIDIA releases Kumo Tabular, a foundation model for tabular prediction

    NVIDIA has released Kumo Tabular, an open-source tabular foundation model that predicts labels for new rows in a single forward pass from labeled table rows, supporting classification and regression without training, tuning, or feature engineering. The model comes in three sizes from 28M to 215M parameters, was pretrained entirely on synthetic tables, is commercially usable under the OpenMDW-1.1 license, and ranks first on four benchmarks: TabArena, BeyondArena, TALENT, and ScoringBench.

    Why it matters: The source reports model sizes, training-data provenance, and rankings on four benchmarks, helping readers assess the accuracy and efficiency tradeoffs of tabular foundation models.

  2. Latent Space78

    AINews: Opus 5.5 excels at generating explainer videos

    Latent Space's AINews roundup of developments from September 24–25 says community feedback was positive after Opus 5.5 launched that week, especially for generating explainer videos with code.

    Why it matters: The weekly roundup offers a quick look at model and agent-infrastructure developments, including community feedback on Opus 5.5's code-generated videos and how vendors position their new models.

9/28Mon
  1. Hugging Face Blog74

    H Company releases Holo4 agent models: 27B dense and 35B-A3B MoE versions

    H Company has released the Holo4 agent-model family, with a 27B dense model and a 35B-A3B MoE model. Both are available through the H Models API, and weights are open on Hugging Face in BF16, FP8, NVFP4, and 4-bit GGUF formats.

    Why it matters: Holo4 opens both its weights and complete evaluation traces, giving readers a basis to assess the cost and capability limits of general-purpose computer-use agents.