Devstral-2-123B vs Gemini Diffusion

Devstral-2-123B

Mistral AI

Devstral 2, released by Mistral AI on December 9, 2025, is a 123 billion parameter dense transformer model specifically designed for software engineering tasks. It features a 256K token context window and achieved 72.2% on SWE-bench Verified at release, making it a competitive open-weight option for automated coding and agentic development. Devstral 2 targets code generation, multi-file software engineering, and agentic development workflows under a modified MIT license.

Gemini Diffusion

Google DeepMind

Gemini Diffusion is an experimental text and code generation model from Google DeepMind, announced at Google I/O in May 2025 as the first diffusion-based language model to achieve quality comparable to autoregressive models on standard benchmarks. Unlike transformer-based models that predict tokens sequentially left-to-right, it generates entire blocks of text by iteratively refining noise — the paradigm used in image and video generation models — enabling faster sampling speeds and stronger mid-generation error correction for code and mathematical editing tasks. At announcement it was available only as an experimental demo via waitlist, with no public API, marking it as a research milestone rather than a production deployment.