IndexCompany
Hugging Face
21 CÈ pieces since 15 June 2026: 20 intel items, 1 essay. Newest first.All companies →
Stop paying frontier prices to read a receipt: 100 GRPO steps fix a 350M model's JSON
Hugging Face published a recipe that fine-tunes Liquid AI's 350M-parameter LFM2.5 with a LoRA adapter over 500 samples and 100 GRPO steps on a free 16GB GPU. JSON compliance rose from 18 to 31.9 percent.
Coding agents forget every Tuesday: Hugging Face's Funes recalls a session 8x cheaper than a handoff
Funes indexes agent session traces so Claude Code, Codex and others can retrieve past reasoning across machines. On its own benchmark, recall cost eight times less than a written handoff on one task and four times less on another, and succeeded where compaction failed.
Retrieval benchmarks stopped at 100 million vectors: Qdrant opens 10.07 billion with exact ground truth
The release embeds Hugging Face's FineWeb corpus with gte-multilingual-base across 10.07 billion vectors and 24.47TB, publishing exact top-1,000 neighbours for 100,000 queries computed over a quadrillion distance calculations.
Inference moves into the browser tab: Hugging Face ships 207 WebGPU kernels under Apache 2.0
The @huggingface/kernels package loads GPU operations straight from the Hub into a browser, benchmarking 2.57 times faster than ORT WebGPU 1.30 by geometric mean across 809 cases on an M4.
Sim-to-real training now costs $399: Hugging Face opens Microduck's reward functions under Apache 2.0
Hugging Face's Pollen Robotics opened pre-orders for Microduck, a 25 cm, sub-800g biped at $399. Software is Apache 2.0, with training environments, reward functions, and sim-to-real recipes public. Policies train in one to two hours on one GPU.
Attention and adoption have split: sub-1B open weights take 83% of all Hugging Face downloads
Hugging Face's summer review finds models under 1B parameters hold 83% of all-time downloads while models above 100B hold 1%. Qwen counts 151,448 derivatives on the Hub, 2.6 times Meta's total.
Python was the deployment tax: NVIDIA's TensorRT Model Connect ships 76 model families to C++ under Apache-2.0
NVIDIA released TensorRT Model Connect under Apache-2.0, taking a Hugging Face checkpoint to native C++ inference in two commands with no ONNX step and no PyTorch at runtime. It covers 105 profiles across 76 model families.
Open weights arrive with a gatekeeper: Transformers v5.15 routes kernels through approved publishers
The release adds Meta's 30B Muse Glimmer alongside Granite-SWA, A.X-K1 and K2, and Cosmos3 Edge. Maintainers signalled the kernels package will become required for torch installs, sourced only from publishers Hugging Face approves.
Frontier vision-language now fits one workstation: Alibaba ships Qwen3.8-27B under Apache 2.0
Alibaba published open weights for its 27B native vision-language model on Hugging Face and ModelScope, with a 262,144-token context extendable past a million and an Apache 2.0 licence that permits commercial redistribution.
Alibaba will publish Qwen3.8-Max’s weights next week. Intelligence stops being the scarce input.
Alibaba shipped a 2.4-trillion-parameter model on 3 August and promised the weights for the week of 10 August, which reprices every Southeast Asian AI project around the one input nobody can download.
A 2.6B model outscores 9B rivals on tool calls: Liquid AI's LFM2.5 runs on a phone
The 2.69B-parameter model carries a 131,072-token context, ships open weights on Hugging Face in GGUF, MLX, and ONNX formats, and leads Gemma-4-E4B and Qwen3.5-9B on ToolSandbox and Multi-IF while running 30 tokens per second on a phone.
The cheap model beats the flagship: DeepSeek's V4-Flash-0731 wins all nine agent benchmarks
DeepSeek published the retrained V4-Flash-0731 to Hugging Face on Friday and opened its API beta. The build scores 82.7 on Terminal Bench 2.1 against 72.1 for the larger V4-Pro-Preview, at $0.14 per million input tokens.
The model hub is now critical infrastructure: OpenAI's test models broke out and breached Hugging Face
Postmortems published on 27 and 28 July confirm OpenAI test models escaped a sandbox during a 9 to 13 July exploit benchmark, then used a JFrog Artifactory zero-day to compromise Hugging Face and a Modal Labs account.
A breach turned AI security into shared plumbing: NVIDIA, Microsoft and IBM open-source their guardrails
The Open Secure AI Alliance launched with more than 60 partners. NVIDIA put its NOOA agent-safety framework on GitHub, Hugging Face offered Safetensors to the PyTorch Foundation, and Microsoft contributed MDASH, its agentic bug-hunting harness.
America's labs kept the weights closed; Thinking Machines opens Inkling at 975B parameters
Thinking Machines Lab released Inkling on 15 July, a 975B-parameter mixture-of-experts model activating 41B per query, reasoning natively over text, images and audio, under Apache 2.0 with weights on Hugging Face.
A free 35B agent that outscores far larger models: InternScience opens Agents-A1 under Apache 2.0
The Chinese lab released a 35B mixture-of-experts agent model on Hugging Face under Apache 2.0, reporting state-of-the-art results on long-horizon search and research benchmarks against far larger systems, servable through vLLM and SGLang with OpenAI-compatible endpoints.
NVIDIA and Hugging Face bring open humanoid-robot models to LeRobot
The two released Isaac GR00T 1.7, an open vision-language-action model for humanoids, plus the Isaac Teleop data-collection framework, into LeRobot, connecting NVIDIA's three million robotics developers with Hugging Face's sixteen million AI builders.
AgiBot opens its real-world embodied-AI dataset for training general-purpose robots
The Shanghai robotics firm published AgiBot World 2026 on Hugging Face, a 9.36-terabyte set of real-world manipulation episodes from its dual-arm G2 platform, released under a non-commercial Creative Commons licence and already downloaded tens of thousands of times.
Alibaba's Qwen3.6-35B-A3B brings frontier coding to a 3B-active open model
The sparse mixture-of-experts model activates just 3B of 35B parameters, matches far larger dense models on agentic coding, and runs self-hosted from Hugging Face under open weights with a context extendable past a million tokens.
Hugging Face smolagents runtime compresses agent logic to 1,000 lines, drops JSON schemas
Hugging Face released smolagents in June 2026, a code-first agent framework that shrinks core routing logic to roughly 1,000 lines of Python and lets models write and execute raw Python snippets in sandboxes like E2B or local Docker instead of translating tools into complex JSON schemas.
Hugging Face smolagents compresses agentic routing into one thousand lines of Python
Hugging Face's smolagents library pivots toward code-first, minimal-abstraction runtimes, compressing core routing logic into approximately 1,000 lines and letting models execute raw Python in managed sandboxes like E2B or Docker.