Today, we are releasing Le Chaton L∃∀N, aka Leanstral 1.5. It achieves SOTA performance on graduate algebra benchmarks FATE-H and FATE-X and improves Pareto Frontier on PutnamBench, solving 587/672 problems with a x10 cheaper budget. 🧵

GLM 5.2 DSpark preview is here! ✨ huggingface.co/RedHatAI/GLM-… This is the first DSpark speculator for a non-DeepSeek frontier model, trained with Speculators and running on vLLM nightly for ~1.5× faster decode for GLM-5.2-FP8 on 4×B300. Stronger checkpoints to come!

💡 Continuous software innovation is the force multiplier behind AI infrastructure — compounding inference performance, lowering cost per token, and increasing long-term value with every optimization. Open source accelerates this advantage. Leading AI frameworks like @PyTorch and inference engines such as @sgl_project and @vllm_project are built natively on NVIDIA CUDA, enabling research breakthroughs and software optimizations to unlock great performance on NVIDIA GPUs from day zero.

Today I deleted PagedAttention from vLLM

🚀 @deepseek_ai's DSpark speculative decoding now runs natively in vLLM! What it is: a semi-autoregressive drafter that proposes several tokens in parallel with non-causal sliding-window attention, then verifies them in a single pass. Output stays identical, decoding takes fewer steps. How vLLM runs it: it reuses the existing SparseMLA backends instead of custom attention kernels, captures the full draft backbone and sampling loop in one CUDA graph, and works with prefix caching and FP8 KV cache. Performance on DeepSeek-V4-Pro-DSpark (verified on NVIDIA 8×B300 GPUs): - ~250 tokens/s at batch size 1 - average acceptance length ~5 - 12-42% higher acceptance than MTP across draft depths Run with vLLM nightly today: vllm serve deepseek-ai/DeepSeek-V4-Pro-DSpark -tp 8 --trust-remote-code --kv-cache-dtype fp8 --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}' DSpark Core PR: github.com/vllm-project/vllm… Thanks @deepseek_ai for open-sourcing DSpark, and to @NVIDIAAI and the vLLM community for landing it! 🙏

🚀 @deepseek_ai's DSpark speculative decoding now runs natively in vLLM! What it is: a semi-autoregressive drafter that proposes several tokens in parallel with non-causal sliding-window attention, then verifies them in a single pass. Output stays identical, decoding takes fewer steps. How vLLM runs it: it reuses the existing SparseMLA backends instead of custom attention kernels, captures the full draft backbone and sampling loop in one CUDA graph, and works with prefix caching and FP8 KV cache. Performance on DeepSeek-V4-Pro-DSpark (verified on NVIDIA 8×B300 GPUs): - ~250 tokens/s at batch size 1 - average acceptance length ~5 - 12-42% higher acceptance than MTP across draft depths Run with vLLM nightly today: vllm serve deepseek-ai/DeepSeek-V4-Pro-DSpark -tp 8 --trust-remote-code --kv-cache-dtype fp8 --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}' DSpark Core PR: github.com/vllm-project/vllm… Thanks @deepseek_ai for open-sourcing DSpark, and to @NVIDIAAI and the vLLM community for landing it! 🙏

Fast, efficient local AI with open-source models just got easier. Qwen3.6-27B-NVFP4 is now on @HuggingFace! It's optimized for NVIDIA Blackwell GPUs & inference ready with @vllm_project. The checkpoint reduces GPU memory requirements by approximately 2.5x for powerful 27B-parameter inference on your own hardware.

During @aiDotEngineer World’s Fair, we’re hosting a happy hour on Thursday, July 2 with special guests from @vllm_project, @inferact, and @buildwithRemy. Join AI builders, founders, and researchers for casual discussion on leading open weight models, inference, infrastructure, and recent developments. Brief intros early on, then drinks and conversation. No conference ticket required, spots are limited!