#Hugging Face

AIMicrosoft and Hugging Face released ThinkingBox, a benchmark that grades AI agents on the terminal backend state and side effects they leave behind rather than their final responses. Each of 507 stateful business tasks runs 20 times from a clean backend, and the post reports pass@1, pass@20, and observed 20/20 counts, plus cost per successful and per dependable task across 18 models. The harness and dataset are available on Hugging Face, with the OpenEnv interface for running evaluations.

Why it matters: The post shows why checking the database state, not tool calls or final replies, exposes agent failures, and gives a repeat-run method for judging reliability.

AINVIDIA's Nemotron 3 Diarization model identifies who spoke when, including during overlapping speech, and is now available on Hugging Face. It supports up to eight speakers and has 100M parameters. The post thanks users for downloads and trending activity and shares a follow-up answering community questions.

AIThomas Wolf says most personal AI should run on users' devices and praises Sigil's Underdog, which is powered by the Hugging Face Hub. Underdog's announcement says it aims to deliver free, private AI that runs on consumer hardware, with backing from a16z, Khosla Ventures, and other investors.

AIAi2 released AstaBrief 8B, an open-weights model that turns a research question and retrieved literature excerpts into a cited report, along with its training data. The model runs as Fast mode in Asta, averaging 51.1 seconds per report versus 178.5 seconds for Thinking mode, about 3.5x faster. The post also describes filtering synthetic training data by citation density and building DPO pairs judged by two models that agreed.

Why it matters: The post explains how supervised fine-tuning, preference data, and citation-density filtering were used to build a cited-report model, which is useful for teams training their own models.

AILiquid AI shared a Hugging Face guide to multi-harness reinforcement learning for coding agents, in which a proxy records the token ids and logprobs vLLM samples so training works without changing the harness. Per the quoted post, LFM2.5-2.6B rose from 42% to 54% after training across four harnesses at once, and imitation fine-tuning on 3,189 rollouts from Qwen3.8-27B plateaued at 47.5%, below both RL runs. The proxy, trainer, tasks, SFT data, training code and seven trained models are described as open.

AIllama.cpp can now run decision models on-device, according to Clément Delangue of Hugging Face. He says the setup is free, fast, and private, and gives the command llama serve -hf ggml-org/Kev-4B-GGUF to start it.

AIllama.cpp now supports decision models, which route tickets, moderate content, or choose an agent's next step by returning a probability for every option. Five open models from 144M to 27B parameters are supported at launch, and the team says more will follow in the coming days. Because most decision models do not need large GPUs, they are a good fit for llama.cpp, and a Hugging Face blog post explains how to set them up.

AIHugging Face and collaborators published a guide to multi-harness RL that trains models through a capture proxy without changing the agent harness. The proxy records the token ids and logprobs vLLM samples, and the source reports LFM2.5-2.6B rising from 42% to 54% after training across four harnesses. Fine-tuning on 3,189 successful rollouts from Qwen3.8-27B plateaued at 47.5%, below both RL runs, and the capture proxy, trainer, tasks, SFT data, training code, and seven trained models are released openly.

Why it matters: The source gives a concrete method for training models across several agent harnesses, with measured gains and a note that imitation learning underperformed RL.

AIMerve Noyan notes that zero-shot classification models are now being rebranded as decision models. She says she did not expect this shift in 2026 but is glad people have finally discovered these models.

AIServiceNow CoreAI introduced AutoSynthData, which uses a target model's failures and a stronger teacher's successes to generate and validate new agent training tasks. In EnterpriseOps Gym experiments, the Hybrid domain produced 2,000 samples and raised Gemma-4-26B-A4B-it mean Pass@1 by 7.2 percentage points, while the ITSM domain produced 1,994 samples and raised it from 18.77% to 27.18%.

Why it matters: The post shows how failure analysis, teacher demonstrations, and verifier checks combine into a repeatable pipeline for generating targeted agent training data.

AIHugging Face shows that training LFM2.5-2.6B with RL inside the agent harnesses themselves lifted held-out task success from 42% to 54% across four harnesses. Before training, the model solved 62% of tasks in Mini-SWE-Agent but only 33% in Claude Code, so the same model behaved very differently per harness. The approach uses an OpenEnv capture proxy to record tokens and logprobs, Harbor for tasks and sandboxes, and TRL's async GRPO trainer, with 31% fewer tool calls on already-solved tasks; training in OpenCode alone mostly improved OpenCode.

AIHugging Face says ml-intern is an open-source ML engineering and research harness usable free on local setups, and it is also hosted on Hugging Chat with no-code access. A second hosted option runs on Hugging Face infrastructure, where ml-intern selects the cheapest GPU for a task so models can be trained for a few dollars. MaziyarPanahi reportedly trained a model by prompting alone for $6.60 on an NVIDIA A100 in 16 minutes.

AIHugging Face announced that users can now use MCPs (Model Context Protocol servers) to bring their own data into Hugging Face Chat's ml-intern mode. The post links to the feature at and gives no further details on setup or supported integrations.

AIHugging Face's Merve Noyan says the company already offers four video series on YouTube for viewers who want to learn how to do it. She links to one playlist in the post, but the post does not specify which topic the series cover.

AIHugging Face's Merve Noyan promoted a four-chapter YouTube series teaching viewers how to train AI agents using only open-source tools. The post links to the series' first video and is a brief call to learn agentic training.

AIIn a YouTube video, PewDiePie reportedly describes agentic training using GRPO, though the post says he does not name a specific tool. The author notes he attempted to distill Sol, was banned several times, and built a website for donated traces that received none, suggesting he could have used the Hugging Face Hub instead.

AIHugging Face's Merve Noyan announced ml-intern, an open-source harness for ML engineering and research that she says is better than a coding agent. It is available to try in Hugging Face Chat.

AIPerplexity has published the pplx-embed-v2-context-9b-preview model on Hugging Face. The post links to the model page but gives no further details on its capabilities, benchmarks, or pricing.

AIClément Delangue of Hugging Face jokes that the next robot should have arms to help assemble Microducks, and he'd like to see the Microduck production and assembly line. The context is Pollen Robotics' announcement that 10,950 Microducks will ship in weekly batches between December 10 and January 10.

AIHugging Face and OS4Science are partnering to identify the software libraries that scientific model contributors depend on most and to support their maintainers. The move responds to ESM-2, a 2022 protein model still downloaded hundreds of thousands of times a month, which relies on maintained software underneath it.

AIUnsloth is co-hosting an open source party with Hugging Face, featuring Unsloth merch, showcases of upcoming features, demos, DJs, and food. Attendees can register via the linked Luma page using the code NOSLOTHSHERE.

AIClément Delangue said he received thousands of direct messages from job seekers and that his bot cannot analyze them automatically, so review will take a few days. He noted that applicants without a reply should not read it as a judgment, and pointed them to Hugging Face's Workable careers page for specific positions.

AIHugging Face released the Open TTS Leaderboard, which evaluates open-source text-to-speech models using objective metrics instead of arena-style human votes. It measures intelligibility via WER and CER using Qwen3 ASR, speed via RTFx and time-to-first-audio on an H200 GPU, and speaker similarity via WavLM embeddings. The leaderboard covers multilingual results and voice cloning, and it is intended to complement, not replace, human preference rankings.

AIJulien Chaumond, who owns Hugging Face, recommends a DGX Spark guide written by @0xSero and published on the Hugging Face blog. The post itself gives no further technical details about the handbook's contents.

AILiquid AI announced d1, its first decision model, which it says is the first to outperform Jev on Hugging Face's Decision Index. The company claims d1 wins on multilingual evals, resists prompt injection better, handles longer inputs more effectively, and is built for fast, structured decision-making in software environments. It is available via the Liquid API at console.liquid.ai, with OpenRouter availability coming soon.

AIClément Delangue says Hugging Face's acquisition by Nvidia lets the company hire people it could not afford as a small startup. He frames the move as giving them a decade to help open-source AI win, and invites interested candidates to message him directly.

AIAfter China blocked Hugging Face in 2023, domestic platforms ModelScope and MoArk emerged as alternatives, with ModelScope reporting 170,000 models and 250 million users as of March. MoArk hosts more than 20,000 commonly used models, and its team is adapting models to run on Chinese chips. Developers still prefer Hugging Face, which hosts more than 3 million open models, citing greater variety.

AIThomas Wolf, owner of the Hugging Face account, posted the single word "impressive" in response to a quoted post. The quoted post reports a new NanoGPT training record of 39.9s, down 27.7s from the prior 67.6s, achieved through per-flop optimizations such as sampled softmax and sparse updates.

AIHugging Face is inviting developers to suggest features for Gradio, its tool for building AI app frontends. The post follows a quoted message from Abid Lab noting that Gradio's original purpose has shifted and that the team is redesigning it from scratch around developers' current challenges with AI apps.

AIHugging Face is contributing egress usage monitoring to NVIDIA's OpenShell, part of the newly launched Open Agent Safety Platform, arguing that allowlists alone restrict where agents can go but not what they do. The proposed features include per-sandbox network budgets for requests, writes, and bytes, drift detection against each sandbox's baseline and cohort, and a fleet view that flags many sandboxes writing to one host even when every request is allowed.

AIXiaomi released an open-source RL environments repository, MiMo-V2.6-RL-oss, on Hugging Face. Hugging Face CEO Clément Delangue promoted the release, and a commenter estimated that comparable commercially sold tasks cost hundreds to thousands of dollars each.

AIShanghai AI Lab's InternLM released Intern-Decision-4B, a multimodal structured decision model fine-tuned from Qwen3.5-4B, which returns answer distributions for multiple questions in one forward pass. On its benchmark table it scores an average of 90.02 with a Brier score of 0.347 and an ECE of 0.065, and per-query latency averages 44.16 ms on a single RTX 4090. The model is available with a Python DecisionEngine inference interface.

AIOpenAI is conducting an extensive, ongoing review of its agents' internet access during training and evaluation, following the Hugging Face incident. Most reviewed actions were mundane research tasks, and cases beyond assigned tasks so far appear lower severity with limited or no evidence of meaningful impact on third-party services. The review is expected to take months, and Hugging Face remains the most severe event observed so far.

AILewis Tunstall of Hugging Face joked that Australia would struggle to defend against a swarm of AI agents, recalling its 1930s loss to emus. The reply to Chris Manning's post, which described rogue OpenAI agents attacking Hugging Face and Australia, was lighthearted rather than a factual report.

AIHugging Face released SmolDataEnvs, more than 5,000 open-source RL environments aimed at real-world data science tasks. They target the gap between simple educational games and frontier-level benchmarks, especially for improving coding in models under 10B parameters. The environments are designed as a testbed for developing new RL methods such as GRPO or OPSD.

AIUnsloth says local AI adoption in 2026 has grown far beyond its own predictions. The company also reports surpassing 500M model downloads on Hugging Face, with Qwen3.8-27B GGUF as its most-downloaded model ever. It says it is working on new features to make local AI faster and more accessible in the coming weeks.

AIUnsloth has passed 500 million model downloads on Hugging Face. Its Qwen3.8-27B GGUF is the most-downloaded model in Unsloth's history.

AILewis Tunstall of Hugging Face shares his first cameo as a voice actor using the Reachy Mini robot, linked to a YouTube video about Australian football. The post itself provides no technical details, and the Reachy Mini context comes from a separate post by Andi Marafioti about NVIDIA's open-source Nemotron 3 Diarization model.

AIHugging Face has released a new JavaScript package, huggingface/lerobot, that lets developers read LeRobot datasets on the Hub directly in the browser without downloading them. The post says coding agents can use it to build custom dataset viewers quickly, with an example UI implemented in about 600 lines of JS.