v0.9.84 v0.27.0
MiniMax-AI/MiniMax-01Aug 5, 2026 → Aug 10, 20265 days between releases
What's New
Upgraded to vLLM v0.27.0 with PyTorch 2.13, Kimi K3 support, and DeepSeek-V4 optimizations.
Breaking Changes
- PyTorch 2.13.0 upgrade (breaking environment change)
- XPU and CPU upgraded to torch 2.13
New Features
- Kimi K3 support with full stack landing
- Qwen3.5 text-only dense and MoE models
- FlashAttention 4 FP8 KV cache support
- Rust frontend gRPC control plane
- DeepSeek-V4 sequence parallelism
Improvements
- DeepSeek-V4 performance push with sequence parallelism
- FlashAttention 4 integration with JIT warmup infrastructure
- Model Runner V2 expansion to non-generative workloads
- First-request compilation stalls removed via runner-owned Triton kernel warmup
- EVS video token pruning for Qwen3.5
Upgrade Notes
Requires upgrading to PyTorch 2.13.0 and torchvision 0.28.0 as this is a breaking environment change. Rust frontend introduces a new gRPC control plane; update configuration accordingly. Verify compatibility with new models like Kimi K3 and Qwen3.5.
The Two Releases
v0.9.84
Aug 5, 2026· 3 months ago
This release brings significant OSINT enhancements including geopolitical analytics, Telegram translation support, ACARS datalink integration, and persistent dashboard layouts.
Full release notesv0.27.0
Aug 10, 2026· 2 months ago
A massive release for the vLLM project featuring Kimi K3 support, Qwen3.5 models, a major PyTorch upgrade, and significant performance optimizations for DeepSeek-V4.
Full release notes