vLLM 官方博客T1·Models
78 pts
Thinking Machines released Inkling 2 Mini, a multimodal MoE with 26B total and 3.5B active parameters, up to 1M token context, and a native DSpark speculator. Its key innovation is a looped architecture where the middle 12 layers run three passes with shared weights, each keeping its own KV cache and reading gated messages from the prior pass. vLLM shipped day-0 support on NVIDIA Blackwell, Hopper, and DGX Spark.