Open-model bonanza: Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 land in one week

This is the densest week of frontier-class open-weights releases since DeepSeek V3 in late 2024. DeepSeek-V4-Flash gets the most attention for its long-context architectural changes — cheaper inference at long contexts is the precise capability open-source has been missing relative to Gemini's 2M window and Claude's 1M. NVIDIA's Kimi-K2.6-NVFP4 quantization, hosted on Hugging Face, makes Moonshot's flagship deployable on NVIDIA inference stacks with FP4 precision — a credibility signal that Nvidia views Chinese open-weights as legitimate inference targets.

PaddleOCR 3.5 landed on a Hugging Face Transformers backend the same day, integrating OCR and document parsing into the standard Transformers pipeline. And Hugging Face's Inference Endpoints v2 (scale-to-zero, agent-aware routing, one-click vLLM) shipped to operationalize all of this.

The Interconnects summary frames it sharply: open weights are abundant and improving fast, but the aggregate-capability gap to GPT-5.5 / Claude 4 / Gemini 2.5 remains. The interesting bet isn't whether open catches frontier — it's whether 'good enough' open-weights at 1/10 the cost reshape the enterprise mid-market faster than frontier labs can move down-market. Cursor Composer 2.5 (built on Kimi K2.5) matching Opus 4.7 and GPT-5.5 on benchmarks at a fraction of the cost is the concrete case study this week.

Watch next: which model lands first on Bedrock as a managed offering, and whether Meta's Llama 4 (rumored post-layoff timing) can reset the open-weights conversation.