AI Coding Live News

ModelScope announced Shanghai AI Laboratory Intern-S2-397B weights in BF16 and FP8 under Apache-2.0. Reported scores include 87.0 on FrontierScience-Olympiad and 84.0 on SWE-bench Multilingual, with long-horizon agent RL, tool use, and thinking/non-thinking modes. Training ingests raw scientific pages (text, visuals, symbols) across 20+ domains. Collection on ModelScope.

Meta Superintelligence Labs launched Muse Voice Transcribe on September 1, its first real-time audio perception model: streaming ASR, diarization for 20+ speakers, and endpointing, with native code-switching, 70+ training languages (25 validated at launch), and hour-plus sessions. Meta says it ranks first on Artificial Analysis streaming speech-to-text as of September 1. It already powers Meta AI for Mac (hold Fn to dictate anywhere) and Muse Code, and is on the Meta Model API at $3 per 1,000 audio-minutes (~$0.18/hour).

Ollama added IBM Granite 4.2 in 3B, 8B, and 30B sizes, Apache 2.0, free for research and commercial use. The family is aimed at enterprise agents: coding, RAG, tool use, structured JSON, multilingual chat, and a built-in thinking mode, with data and training run through IBM GRC and document-quality reviews. Context is 128K; ollama run granite4.2 is enough to start locally. For teams that cannot send code off-network but still want a reason-on/off small model with a commercial license, it is a compliance-friendly local agent base.

GitHub said dictation in VS Code now uses a multilingual on-device model by default, keeping audio on the machine. It can follow the configured language, the system or browser locale, or auto-detect speech. For developers who talk Copilot through edits, comments, or agent prompts, ASR no longer has to leave the device on the default path. That cuts both latency and the audio exfiltration surface for code-adjacent speech.

The SWE-bench benchmark team launched SWE-bench Pro, expanding from Python to Rust, Go, C++, and TypeScript. Featuring 1,200 curated real-world issues from high-traffic open-source repositories, it provides a unified industry standard for multi-language autonomous agent software engineering evaluation.

Mistral AI rolled out a major update to Codestral 2501, expanding native context length to 256K tokens. The model achieves over 98% accuracy on 200k Fill-in-the-Middle (FIM) benchmarks, with broad multilingual fluency spanning 80+ mainstream and emerging languages including Rust, Go, Python, Solidity, and Zig.

Google DeepMind released Gemini 3.5 Transcribe, an intelligent speech-to-text model designed for clean and structured audio transcription across 85+ languages, now live in Antigravity and AI Studio.

Cohere added North Micro Vision to its open-source family under Apache 2.0. At 2.4B parameters, it supports native-resolution vision, multi-turn multimodal chat, spatial grounding, and multilingual image understanding, outperforming Gemma 4 E2B and Ministral 3 3B on document VQA.

Aider creator Paul Gauthier announced that OpenAI o3-pro with high reasoning effort set a new record of 85% on the Aider Polyglot coding benchmark. The benchmark evaluates multi-language real-world codebase editing, solidifying o3-pro's position at the top of practical coding leaderboards.