Daily News · 2 min read

Hugging Face AI Updates: June 10, 2026

hugging-face models code-generation speech-recognition ai-agents

1. Cohere Labs Publishes North Mini Code, a 30B Mixture-of-Experts Model for Developers

Hugging Face. Cohere Labs published North Mini Code on the Hugging Face blog, describing it as the company’s first model aimed at developers and agentic software engineering tasks. The model is a 30-billion-parameter Mixture-of-Experts decoder-only Transformer with 3 billion active parameters and 128 experts, 8 of which activate per token, trained with supervised fine-tuning followed by reinforcement learning with verifiable rewards. It scored 33.4 on Artificial Analysis’ Coding Index and is available in OpenCode, through the Cohere API, and on Hugging Face in BF16 and FP8 formats under the Apache 2.0 license. Source

2. ServiceNow-AI Benchmarks Frontier ASR Systems on Code-Switched Speech

Hugging Face. A ServiceNow-AI post on the Hugging Face blog benchmarked seven automatic speech recognition systems on code-switched speech, where bilingual speakers blend two languages within a single utterance. The evaluation used 918 synthetic utterances across Spanish-English, French-English, Canadian French-English, and German-English pairs, testing systems including ElevenLabs Scribe V2, AssemblyAI Universal 3-Pro, Google Gemini 3 Flash, Deepgram Nova 3 Multilang, Mistral Voxtral Small 24B, Nvidia Parakeet TDT 0.6b V3, and OpenAI Whisper Large V3 Turbo. The authors reported that Scribe V2, Gemini 3 Flash, and AssemblyAI Universal 3-Pro were the top performers, that code-switching imposed minimal penalties on the strongest systems, and that errors concentrated on embedded English segments. Source

3. Community Post Shows an Agent Chaining Two Spaces to Build a 3D Gallery

Hugging Face. A community article by Mishig Davaadorj on the Hugging Face blog described an AI agent that built a 3D Paris gallery of monuments rendered as interactive Gaussian splats by chaining two Hugging Face Spaces in sequence. The agent connected the ideogram-ai/ideogram4 image generation Space to the VAST-AI/TripoSplat single-image 3D reconstruction Space without hardcoded integrations. The post highlights an agents.md file exposed by every Gradio Space that provides the schema URL, call and poll templates, file upload instructions, and an auth hint, allowing agents to call Spaces programmatically without SDKs. Source