Open-model release wave (DeepSeek V4, Vision, Gemma 4, GLM-5.2, MiniMax M3, Kimi K2.7, MiMo-V2.5, Qwen3-Coder-Next, DiffusionGemma, and more)
Key Questions
Which major open-weight models were released recently?
Releases include MiniMax M3 with 1M context, GLM-5.2 (753B MoE), Kimi K2.7 Code (1T MoE), MiMo-V2.5 (310B MoE), DeepSeek V4 with DSpark, Gemma 4 QAT, DiffusionGemma, and Qwen3-Coder-Next.
How does MiniMax M3 compare to other models?
MiniMax M3 beats Opus 4.7 and GPT-5.5 at 50x lower cost with its 1M context length. It is part of the wave of Chinese labs dominating top open-source models.
What is notable about Kimi K2.7 Code in GitHub Copilot?
Kimi K2.7 Code is the first open-weight model from a PRC lab integrated into GitHub Copilot for enterprise coding. It offers lower costs with different audit considerations.
Where can MiMo-V2.5 be accessed?
MiMo-V2.5 from Xiaomi is available on DeepInfra. It unifies agentic and multimodal capabilities with 1M context and 15B active parameters under MIT license.
What efficiency features does DeepSeek V4 include?
DeepSeek V4 features DSpark for 60-85% faster inference and DeepSpec for custom draft models, both under MIT license. It integrates with Red Hat via GLM-5.2.
How do open-weight models affect GitHub Copilot?
GitHub now supports open-weight models like Kimi K2.7 Code in Copilot for the first time. This enables lower-cost options while introducing new data law considerations for Chinese models.
What is OpenRouter's role in the model landscape?
OpenRouter fuses budget models to outperform GPT-5.5 at half the cost. This highlights the growing competitiveness of open-weight releases.
Why are Chinese labs prominent in these releases?
Chinese labs released GLM-5.2, Kimi K2.7, MiMo-V2.5, and others that lead in coding, security, and multimodal benchmarks. They broaden capabilities while raising considerations around data laws and audits.
Continuing wave of major open-weight releases: MiniMax M3 (1M context, beats Opus 4.7/GPT-5.5 at 50x cost), GLM-5.2 (753B MoE, MIT, #1 frontend coding, security implications), Kimi K2.7 Code (1T MoE, 32B active) now in GitHub Copilot as first open-weight model in enterprise coding assistant, MiMo-V2.5 (Xiaomi, 310B MoE, 15B active, MIT, unified agentic+multimodal, 1M context, on DeepInfra), DeepSeek V4 with DSpark (60-85% faster inference, MIT) and DeepSpec for custom draft models, Red Hat integration with GLM-5.2. Also DeepSeek Vision, Gemma 4 QAT, DiffusionGemma, Qwen3-Coder-Next, etc. Chinese labs dominate top open-source models. OpenRouter fusion of budget models beats GPT-5.5 at half cost.