Today for AI
HOT RADAR
llmHEAT 13.1°

AI CURATED HOT STORY · 10/29/2026

vLLM Adds Day-0 Support for Mistral Large 4 with Sparse Attention and Multimodal Features

1 reports archived1 independent sourcesupdated 10/29/2026, 08:00:00
Synthesis & Latest Updates
1 Sources Cross-Validated

vLLM announces Day-0 support for Mistral Large 4, a frontier MoE native multimodal model. The integration features a DeepSeek-style sparse attention stack, cross-layer KV sharing, NoPE layers, and built-in speculative decoding, driving generic improvements in context parallelism and host-memory KV offload within the vLLM framework.

LATEST/vLLM announces Day-0 support for Mistral Large 4, a frontier MoE native multimodal model. The integration features a DeepSeek-style sparse attention stack, cross-layer KV sharing, NoPE layers, and built-in speculative decoding, driving generic improvements in context parallelism and host-memory KV offload within the vLLM framework.

Hourly heat curve

3 fully observed hours
Now
13.1
Peak
13.810-30
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first
  1. vLLM 官方博客T1·88 pts

    vLLM Adds Day-0 Support for Mistral Large 4 with Sparse Attention and Multimodal Features

    • Mistral Large 4 uses a 72-layer MoE architecture (1T total params, ~49B active), combining MLA with a DSA sparse indexer.
    • Features include cross-layer KV sharing, NoPE layers, SSMax/LeRoPE, and a built-in DSpark speculative decoder with Pixtral vision encoder.
    • Implemented on vLLM's Model Runner V2 (MRV2) using breakable CUDA graphs, optimizing decode/prefill context parallelism and HiSparse KV offload.
vLLM Adds Day-0 Support for Mistral Large 4 with Sparse Attention and Multimodal Features | Today for AI