Hugging Face

Hugging Face released @huggingface/kernels alongside 207 WebGPU kernels on the Hub, each packaged with its interface, shader templates, correctness cases and benchmarks. Fleet, an in-browser benchmarking suite, scores them on your own GPU and feeds results back to the community.

Tools for the Reachy Mini conversation app can now live in public Hugging Face Spaces and be called over MCP — added with one command, executed remotely, nothing downloaded — while each profile's tools.txt decides which capabilities the model is allowed to reach for, alongside the local body tools.

Qwen releases a text-to-image prompt-rewriting model that turns short, any-language requests into detailed English prompts with a recommended aspect ratio, feeding into Qwen-Image-2.1.

NVIDIA released the BF16 Nemotron 3.5 Super VL 120B A12B, a multilingual vision-language MoE tuned for reasoning and agent use, with the post-training dataset listed and vLLM and SGLang serving supported.

Microsoft and Hugging Face turn ThinkingBox's database-state grading into RL rewards. GRPO on Qwen3.8-27B lifts average success from 51.7% to 60.8% on tasks outside the benchmark, with training code released.

SkyPilot adds Hugging Face storage as a first-class backend. A store: hf file mount pulls a Bucket read-write or any model, dataset or Space repo read-only into a job, so training and serving can land on whichever of 20+ clouds has GPUs, with no egress and Xet-level dedup on checkpoints.

How huggingface_hub went from 4–6 week releases to weekly inside one GitHub Actions workflow, using open-weight models for the judgment work — release notes and announcements — with grounding against fabrication and a single human review checkpoint.

Alibaba's Qwen team open-sources Qwen-Image-2.1, a 7B-parameter unified text-to-image and editing model with native RGBA transparency support, up to 10 reference images for edits, and mask/annotation-based local editing.

IBM Research scaled its ALTK-Evolve memory system across eight models, from a 30B dense model to frontier systems, and found that agentic memory behaves like a dose rather than a switch: how many distilled guidelines you inject, and how, decides both accuracy and cost.

NVIDIA's NeMo AutoModel subclasses Transformers v5's AutoModelForCausalLM and adds expert parallelism, DeepEP fused all-to-all dispatch and TransformerEngine kernels, reporting 3.4-3.7x fine-tuning throughput and 29-32% less GPU memory on MoE models from 30B up to a 550B multi-node run.

Ukisai post-trained Qwen3.8-27B to suppress tokens linked to reasoning 'overthinking' rather than just truncating chain-of-length, then recovered accuracy via on-policy distillation, yielding claimed 58% shorter outputs and ~2x speedup, with open weights, GGUF quants and a rate-limited free API.

InclusionAI's Ling-3.0-flash-VL adds native image/video understanding to its 124B/5.5B-active MoE, wiring vision into agentic reasoning and interface actions with a 1M-token context and a measured intelligence-index gain over the text-only model.

Ai2's Skylight team walks through Shippy, a maritime analysis agent: behaviour split into soul, skills and config; skills written in the same markdown spec coding agents use; deterministic tool calls that resolve regions before querying; sandboxed hosting; and evals aimed at the agent, not the model

Meta's model organization published an NVFP4 quantized checkpoint of Muse Spark Open 816B A42B, a multimodal mixture-of-experts model, for lower-memory serving through Transformers and vLLM.

Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights model for text-to-image and editing with up to 10 reference images and a Qwen3-VL-4B text encoder. It performs poorly on short prompts without a bundled prompt upsampler.

IBM's time series foundation models are in Early Access on Confluent Cloud, with Confluent Platform to follow. The pitch is stream-native inference: one model trained across many signals generalizes to series it has never seen, replacing bespoke per-series models and the safety margins teams use ins

Ai2 released OlmoEarth v1.3, an embedding model over Sentinel-1, Sentinel-2 and Landsat time series, plus precomputed global annual embeddings (64 and 128 dimensions) and a no-sign-in explorer for change tracking and finding similar places.

Liquid AI released DSpark, a speculative-decoding draft model for LFM2.5-VL-3B that delivers up to 2.66x decode speedup on H100 GPUs and over 3x on Apple silicon, with published draft-acceptance and throughput benchmarks.

Together releases an 8B agent checkpoint on Qwen3-8B-Base that switches between latent reasoning and text tool use over shared memory. Latent thoughts are readable as pseudo-text; code and MIT weights are public.

Hugging Face rebuilt hf for two audiences: it reads CLAUDECODE, CODEX_SANDBOX and similar variables, then renders un-truncated, un-colored, token-lean output with next-command hints and retry-safe semantics. Benchmarks against Claude Code and Codex put the no-CLI baseline at up to 6x the tokens.

OpenEnv moves to committee governance spanning Meta-PyTorch, Nvidia, Modal, Prime Intellect, Unsloth, Microsoft and others, and tightens its remit: a standard for publishing, deploying and consuming RL environments, explicitly not rewards or training loops.

Hugging Face's biannual open-model report: 2.96M public model repos where 1.5% take 99.2% of downloads, Chinese labs skipping the small-to-large progression with 754B–2.78T ceilings while most US open releases stayed under 130B, and AMD and NVIDIA leading new-repo counts.

The 4-step distilled Turbo variant of InclusionAI's open LLaDA-Image model uses Twin-DMD distillation for fast generation and editing, released alongside the Base checkpoint with Diffusers inference code.

NVIDIA publishes an NVFP4-quantized build of Qwen3.8-Flash-Next, reporting accuracy within roughly a point of the FP8 baseline across GPQA Diamond, SciCode, and other benchmarks while cutting the memory and compute footprint for agentic deployments.