AI news — February 2026
Every AI story covered in February 2026, summarized. From the @rotative_ai daily feed.
stories
- Hugging Face adds first-class Mixture-of-Experts support
hugging face refactored transformers to support mixture-of-experts (moe) models natively. the update addresses weight loading, device placement, and distributed execution for sparse architectures like deepseek r1 and mixtral. this allows models to use fewer active parameters during inference while maintaining total capacity, significantly improving compute efficiency and speed.
- Hugging Face and Unsloth offer free credits for model fine-tuning
hugging face and unsloth partnered to offer free compute credits and a one-month pro subscription for fine-tuning small language models via hugging face jobs. users can now deploy coding agents like claude code to automate training, script generation, and monitoring directly from the cli. the program targets lightweight models like lfm2.5-1.2b-instruct.
- llama.cpp creator Georgi Gerganov joins Hugging Face
ggml and llama.cpp creator georgi gerganov joined hugging face to scale local ai infrastructure. the project remains 100% open-source with full technical autonomy for gerganov's team. the partnership aims to integrate the transformers library with llama.cpp for seamless local deployment, targeting ubiquitous local model access.
- IBM and UC Berkeley use MAST to diagnose IT agent failures
ibm and uc berkeley applied the mast taxonomy to 310 itbench traces involving gemini-3-flash, kimi-k2, and gpt-oss-120b. the analysis revealed that frontier models fail cleanly while open models suffer cascading failures. incorrect verification predicts failure most often across all models.
- Hugging Face Gradio enables custom frontend components in single Python files
hugging face gradio now supports complex custom frontends via a single gr.html subclass, eliminating separate build steps. developers can now integrate three.js 3d controls, real-time transcription displays, or native drag-and-drop kanban boards directly into python pipelines without react.
- Hugging Face builds agent skill for writing production CUDA kernels
hugging face released a new agent skill that enables coding models like claude and codex to write production cuda kernels. the agents successfully generated working kernels for diffusers and transformers, complete with pytorch bindings and benchmarks. the 550-token skill provides gpu architecture awareness for h100, a100, and t4 chips and handles integration patterns.
- OpenEnv framework evaluates AI agents against real-world APIs
meta and hugging face released openenv, a framework using real environments instead of simulations to test ai agents. turing contributed a production-grade calendar benchmark which revealed agents drop to 40% success when handling natural language ambiguity and struggle with multi-step reasoning.
- Transformers.js v4 launches with C++ WebGPU runtime and 4x BERT speedup
huggingface released transformers.js v4, featuring a rewritten c++ webgpu runtime enabling local inference in node, bun, and deno. the update introduces a monorepo structure, splits the 8,000-line models file into modules, and adds support for gpt-oss, mamba, and mixture of experts architectures. benchmarking shows a 4x speedup for bert-based embedding models.
- ServiceNow releases SyGra Studio for visual AI workflow orchestration
servicenow released sygra studio, a visual interface for designing and validating ai workflows. the tool supports openai, azure, and hugging face endpoints with real-time debugging and cost tracking. users can build complex pipelines using llm nodes and structured outputs. documentation is available at servicenow.github.io.
- Hugging Face introduces decentralized community evaluations
hugging face is decentralizing ai evals. dataset repos now host leaderboards via eval.yaml. users submit results via pull requests with .eval_results/*.yaml. verified badges ensure reproducibility. this creates a single source of truth for benchmarks like mmlu-pro and gpqa.
- H Company launches Holo2-235B-A22B for UI localization
h company released holo2-235b-a22b preview, a 235b parameter model for ui element localization. it hits 78.5% on screenspot-pro and 79.0% on osworld g, setting a new sota record. the model uses agentic localization to iteratively refine predictions, improving accuracy by 10-20%.