←Search results

Search “Hugging Face”

By relevance · Total 1210 results Clear search

Sunday

  1. Story1 reports · 1 facts

    Fireworks publishes GLM 5.2 numerical alignment experiment

    On September 30, 2026, Fireworks published a GLM 5.2 numerical alignment experiment. The experiment shows that when running the same reinforcement learning task on GLM 5.2, an inconsistency between the trainer and the rollout engine in how token probabilities are computed leads to a KL of about 0.013, roughly 45% of samples discarded per batch, and the reward dropping from about 0.9 to below 0.2 around step 20; after numerical alignment, the KL is 0 and the reward remains stable throughout. The same material also documents a Qwen3.5-MoE experiment: precision differences in the weighted sum of expert outputs brought the k3 of image tokens to 0.296, about 300 times the equivalence threshold, and k3 dropped to zero after merely replacing the MoE module with the Hugging Face reference implementation. Currently, this experiment is published by Fireworks as a first-party source, and no subsequent independent progress has emerged.

    ent, the KL is 0 and the reward remains stable throughout. The same material also documents a Qwen3.5-MoE experiment: precision differences in the weighted sum of expert outputs brought the k3 of image tokens to 0.296, about 300 times the equivalence threshold, and k3 dropped to zero after merely replacing the MoE module with the Hugging Face reference implementation. Currently, this experiment is published by Fireworks as a first-party source, and no subsequent independent progress has emerged.

    Matched fieldsSummary

  2. Story1 reports · 1 facts

    StepFun and ACE Studio launch the StepAudio 3 Music music generation model

    StepFun and ACE Studio jointly released StepAudio 3 Music, a foundational music generation model. This is StepFun's first foundational music generation model, capable of generating complete songs based on prompts and lyrics. The model integrates four workflows: song generation, accompaniment generation, music cover, and vocal-to-song arrangement. ABC-COT plans the musical structure and arrangement before synthesis, and users can rewrite prompts, edit ABC notation, and iterate. At the time of release, a Hugging Face Space demo and API documentation were already available, and the ABC-COT API is coming soon. There are currently no reports of further progress.

    e of generating complete songs based on prompts and lyrics. The model integrates four workflows: song generation, accompaniment generation, music cover, and vocal-to-song arrangement. ABC-COT plans the musical structure and arrangement before synthesis, and users can rewrite prompts, edit ABC notation, and iterate. At the time of release, a Hugging Face Space demo and API documentation were already available, and the ABC-COT API is coming soon. There are currently no reports of further progress.

    Matched fieldsSummary

  3. Story1 reports · 1 facts

    Bailing releases and open-sources native multimodal model Ling-3.0-flash-VL

    On September 9, 2026, Ant Bailing released and open-sourced Ling-3.0-flash-VL, the first native multimodal large model in the Ling series. Built as an extension of the Ling-3.0-flash MoE architecture, the model has 124B total parameters and 5.5B activated per inference, natively supports image, text, and video inputs, and has a 256K-token context window. According to the company, native multimodal joint training improved the model's score on the Artificial Analysis Intelligence Index v4.1.1 by 4 points over the text-only version, and it introduces a visual feedback loop of observe, act, verify, and correct, used for front-end replication, GUI agents, and medical report interpretation. The model is currently available for free trial on Ling Studio, and the BF16 and FP8 versions have been open-sourced on Hugging Face and ModelScope.

    ontext window. According to the company, native multimodal joint training improved the model's score on the Artificial Analysis Intelligence Index v4.1.1 by 4 points over the text-only version, and it introduces a visual feedback loop of observe, act, verify, and correct, used for front-end replication, GUI agents, and medical report interpretation. The model is currently available for free trial on Ling Studio, and the BF16 and FP8 versions have been open-sourced on Hugging Face and ModelScope.

    Matched fieldsSummary

Tuesday

Sunday

  7. Story9 reports · 2 facts

    Qwen-Image-2.1 supports ComfyUI

    On September 20, 2026, the Qwen team open-sourced the Qwen-Image-2.1 image model, unifying text-to-image and image editing into a single model that supports a variety of image editing tasks. According to the official introduction, the model's visual generation component is a 7B-parameter, 32-layer Single-Stream DiT that natively supports transparent image generation and editing, supports up to 10 reference images as well as local editing methods such as circling and masking, and reduces inference overhead and VRAM usage during multi-image editing through mixed-granularity attention and KV cache reuse; the model weights have been released on GitHub, Hugging Face, and ModelScope. On the same day, Qwen announced that Qwen-Image-2.1 is natively supported in ComfyUI; users can update ComfyUI or use Comfy Cloud, download the weights from Hugging Face, and load the template to use it.

    up to 10 reference images as well as local editing methods such as circling and masking, and reduces inference overhead and VRAM usage during multi-image editing through mixed-granularity attention and KV cache reuse; the model weights have been released on GitHub, Hugging Face, and ModelScope. On the same day, Qwen announced that Qwen-Image-2.1 is natively supported in ComfyUI; users can update ComfyUI or use Comfy Cloud, download the weights from Hugging Face, and load the template to use it.

    Matched fieldsSummary

Wednesday

Wednesday

Friday

  1. Story1 reports · 1 facts

    Baseten and Base Labs Announce Building Open Infrastructure for Model Safety Evaluation and Monitoring

    On September 16, 2026, Baseten and Base Labs announced a joint effort with Goodfire and Hugging Face to build safety evaluation and monitoring infrastructure for the open-source models they serve, integrating safety research directly into training and inference systems. The work covers three areas: the training phase, where models are made to follow explicit policies, recognize the boundaries of their capabilities, and proactively seek help; the monitoring and intervention phase, which studies inputs, outputs, tool calls, and models' internal states, and connects signals to permissions, isolation, and pre-execution checks; and learning from failures. In terms of division of labor, Base Labs is responsible for developing and publicly releasing the methods, while Baseten integrates them into its deployment infrastructure, offering them as a hosted service and publishing the methods and evidence so that customers and external evaluators can verify them independently. There are currently no further developments.

    On September 16, 2026, Baseten and Base Labs announced a joint effort with Goodfire and Hugging Face to build safety evaluation and monitoring infrastructure for the open-source models they serve, integrating safety research directly into training and inference systems. The work covers three areas: the training phase, where models are made to follow explicit policies, recognize the boundaries of their capabilities, and proactively seek help; the monitoring and intervention phase, which studies i

    Matched fieldsSummary