Open-weight models and the AI price/performance war
Key Questions
What is DeepSeek V4's pricing and performance?
DeepSeek made V4 Pro permanently available at a 75% discount, matching GPT-5.5 Pro at 1/200th the cost. The company is raising $7.4B in new funding.
How does Kimi K2.6 perform compared to leading models?
Kimi K2.6 outperforms Opus and Gemini on frontend tasks at 94% lower cost. Kimi K2.7-Code is also releasing with improved token efficiency.
What open-weight models were recently released?
MiniMax M3 offers 1M context and beats GPT-5.5 on coding benchmarks. NVIDIA released open models Cosmos 3 and Nemotron 3 Ultra.
How does GLM-5.2 compare on benchmarks?
GLM-5.2 from Zhipu AI beats GPT-5.5 on SWE-bench Pro at one-sixth the cost. It emphasizes coding and agentic improvements with a 1M-token context.
What funding rounds are rumored for European AI companies?
Mistral is rumored to raise €3B at a €20B valuation. This would nearly double its previous valuation.
What community trends are emerging around local models?
Interest is growing in local open-source models for coding tasks. Benchmarks are evaluating open-source LLM agents for SQL generation.
What is Xiaomi contributing to the open-source space?
Xiaomi released MiMo Code V0.1, an open model focused on coding. It joins other releases emphasizing cost efficiency and performance.
How is the AI price war affecting the market?
Cost advantages from GLM-5.2 and DeepSeek are intensifying competition. This puts pressure on closed models from OpenAI and Anthropic.
Alibaba’s Qwen3.8-Max 0902 snapshot adds a major long-context coding release, with a reported CodeArena score of 1,691 that still needs independent validation. It joins DeepSeek V4, GLM-5.2, MiniMax M3, Kimi K2.7-Code, Gemma 4, Mistral, and other open-weight models driving down inference costs and expanding self-hosted and local deployment options.