Hugging Face
1 Tool · 170 Intel
Hugging Face is an American machine learning company, founded in 2016 in New York City by Clément Delangue, Julien Chaumond, and Thomas Wolf. It began as a teen-focused chatbot app before pivoting to become a widely used open-source platform and community for sharing machine learning models, datasets, and tools, including its popular Transformers library. It is often described as functioning like a "GitHub of machine learning" for researchers, data scientists, and engineers.
Is this you? Sign in with X to claim this profile.
Tools
Open-source ML-engineer agent from Hugging Face that reads research papers, trains models, and ships them end to end.
Intel
Xiaomi released MiMo-V2.6-Flash-RL, a faster multimodal sibling to Pro-RL, open-sourced under MIT with vision, audio, and video understanding for lower-latency multimodal inference.
The Swiss AI Initiative details Apertus 1.5: continued pretraining on 2 to 4 trillion more tokens, image and audio via tokenizer-level early fusion, a thinking mode, 262k context and improved tool use, all with open weights, data and training details.
Liquid AI releases LFM2.5-24B-A2B, a reasoning-only MoE model with 2B active parameters, 131K context and tool calling. It reports strong agentic benchmark scores and fast CPU and GPU inference for on-device agents.
InclusionAI released LLaDA2.2-mini, a lightweight diffusion LLM using Levenshtein Editing (DELETE/INSERT control tokens) to self-correct output during generation, targeting long-context tool calling and multi-turn agent interaction, with out-of-box vLLM and SGLang serving.
Hugging Face explains how tokenizers v1 achieves encode/decode speedups often exceeding 10x over v0.23, aimed at keeping GPUs fed during large-scale LLM training and serving.
Hugging Face ships AnyLanguageModel 1.0 alongside swift-huggingface and swift-xet 1.0. It keeps Apple's Foundation Models API while routing to the system model, llama.cpp, MLX, Core ML or hosted providers, with tool calling and guided generation across all of them.
Xiaomi released MiMo-V2.6-Pro-RL, an open MIT-licensed multimodal model (vision, audio, video, long context, agent-tagged) with day-one vLLM and SGLang serving support.
Xiaomi's MiMo team shipped MiMo-V2.6-Distill-Qwen-9B, a 9B image-text-to-text model distilled from Qwen3.5-9B and fine-tuned for agentic, tool-use, and coding tasks, ready to serve via vLLM, SGLang, or Transformers.
Liquid AI ships DSpark draft models for LFM2.5-1.2B, 2.6B, and 8B-A1B, adding speculative decoding for up to 3.2x faster inference and 57% lower function-calling latency at ~300M parameters, with quality unchanged under greedy decoding and upstream llama.cpp and SGLang support.
AWS details Strands Robots: one agent loop that records robot demonstrations, keeps them in LeRobot format inside Xet-backed Hugging Face Storage Buckets, trains by streaming straight from the Hub, and pushes the policy back to hardware — with byte-level dedup to stop paying for the same transfers.
Gradio's gr.Workflow turns the pipeline itself into the UI: describe steps as a graph of typed nodes and get a drag-and-drop canvas where each node runs independently and every intermediate result is inspectable. The same graph doubles as a REST API and deploys to Spaces in one command.
TII releases Falcon-Emirati-7B, adapted from its Falcon-H1-Arabic hybrid SSM-attention model using web, MSA-culture and synthetic glossary-guided data, evaluated by native speakers, automatic benchmarks and pairwise LLM judging.
NVIDIA released an NVFP4-quantized version of DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with 1M-token context, with vLLM and SGLang deployment support tested on Blackwell GPUs for cheaper self-hosted inference.
IBM Research's ScarfBench evaluates coding agents on cross-framework Java migrations between Spring, Jakarta EE and Quarkus — 34 applications, 204 tasks, ~151K lines — scored on whether the migrated app builds, deploys and passes 1,331 expert-written tests.
Agentic Resource Discovery is a draft open spec, with contributors from Microsoft, Google, GoDaddy and Hugging Face, that adds a discovery layer in front of MCP, Skills and A2A: an ai-catalog.json manifest plus a POST /search registry API for runtime capability lookup.
Ant Group's first finance-enhanced Ling model, continued-trained on financial data with 124B total/5.1B active parameters and 256K context, evaluated on agentic finance benchmarks like FinFIRST and SpreadsheetBench for research and modeling workflows.
Meta released Muse Spark Open 816B A42B weights on Hugging Face, a multimodal MoE built for agentic coding, with Transformers, vLLM and Docker usage documented. It ships under a custom Muse model license.
Mistral released Voxtral Mini 4B Realtime Arabic, a 4.4B streaming speech-to-text model for Arabic dialects and Modern Standard Arabic, reaching 8.82% average CER at 480 ms delay and served through vLLM or Transformers.
Nex AGI's mid-tier Nex-N2.5-Pro rounds out its new agentic model family alongside mini and Max, continuing Nex-N2's computer-use and browsing focus; weights are listed as forthcoming rather than released.
NVIDIA published a LiLiCorr speculative-decoding draft checkpoint for the Nemotron 3.5 Bolt 74B A9B model. It proposes tokens that the target model verifies, under the OpenMDW 1.1 license, for lower-latency serving.
InclusionAI shipped SingProbe for Ling-3.0-flash: a 5.18M-parameter probe reading hidden states at layers 13, 26 and 40 to emit per-token intent, unsafety and hallucination scores. MIT licensed, wired into SGLang and vLLM branches, 0.03% false positives on benign responses.
Black Forest Labs' FLUX.3 Klein lands in Diffusers with text-to-image, multi-reference editing and bounding-box layout control. Covers distilled and base checkpoints, quantized weights, LoRA training scripts and the modular pipeline.
A public, low-cost GRPO recipe using TRL that takes LFM2.5-350M from 22.6% to 29.7% on the IFStruct structured-output benchmark in about 500 samples and 100 steps. Covers reward design, LoRA, merging, and local evaluation through llama.cpp.