At @NVIDIAAI we continue to push open data, techniques and models forward because we know that every organization needs the freedom to build and deploy AI in their own way. We're now the biggest institutional contributor on HuggingFace and we expect to continue publishing. It's not charity or a science project - we know that when AI grows, NVIDIA's opportunities also grow. More analysis on the state of open source AI here:

Benchmarking @NVIDIAAI's Nemotron Puzzle 75B locally on the GX10. NVFP4 via vLLM's OpenAI API, MTP speculative decoding, forced 1,500-token generations. 🏃‍♀️22.75 tok/s in a single session into 88.85 cumulative at 7 sessions. 🧍Baseline without MTP: ~16.3. Scripts + setup:

I tested @NVIDIAAI's Nemotron Puzzle 75B locally on my GX10 with 3 HTML game challenges: - Tower Defense 🗼(~ 1m30 at 60 tok/s) - Wrecking ball 🏐(~ 40 s at 60 tok/s) - 2D Roguelite 👾(~ 1m at 60 tok/s) The results are satisfying considering that there was no harness involved (ignoring the shaggy physics of the wrecking ball one)

i added native support for @NVIDIAAI's Nemotron Puzzle 75B to mlx-lm. it now runs natively on an M2 Max 64GB: ⚡️22 tok/s 💾45.5 GB peak memory usage 📚4-bit experts + 6-bit dense + BF16 head i also fixed an annoying numerical bug in mlx-lm. outputs were subtly wrong, cosine similarity was 0.8832 vs NVIDIA's reference (identical inputs). the culprit was one dtype cast in the Mamba layers happening in a different spot than NVIDIA's. once moved, the cosine similarity improved to a satisfying level (0.999...). related PR: weights:

RL fine-tuning is now live for @nvidiaai Nemotron 3 on Fireworks, starting with Nemotron 3 Super (LoRA). Train with GRPO and serve the model in one place. We price by GPU-hour, not per token, so long multi-turn rollouts don't blow up your bill. Training shapes →

Red Hat AI just shipped DFlash speculator checkpoints for two of @NVIDIAAI's most powerful open models: → Nemotron Ultra 550B → Nemotron Super 120B On math and reasoning: ~5 out of 7 draft tokens accepted on average. On code (HumanEval): ~3.4 out of 7. Both checkpoints trained with the open source Speculators library from @vllm_project. Apache 2.0. Validated on NVIDIA B200. One flag to enable in vLLM: --spec-model RedHatAI/NVIDIA-Nemotron-3-Ultra-550B-A55B-speculator.dflash --spec-tokens 7 --spec-method dflash 🔗 Ultra 550B: 🔗 Super 120B:

🚀 @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: Thanks @deepseek_ai for open-sourcing DSpark, and to @NVIDIAAI and the vLLM community for landing it! 🙏

We’re excited to share that @nvidia Nemotron 3 Nano Omni is now available at launch on Bitdeer AI Model Studio 🎉. As part of the NVIDIA #Nemotron# 3 family, Nemotron 3 Nano Omni is an open multimodal model that unifies reasoning across video, audio, images, and text， simplifying agentic AI development with leading efficiency and accuracy. From multimodal understanding to real-world enterprise workflows, Nemotron 3 Nano Omni helps reduce the complexity of fragmented model stacks while enabling richer cross-modal context for search, analysis, summarization, and automation. Access and run it today on Bitdeer AI Model Studio through a simple serverless #inference# API, making it easier to move from experimentation to production. 🔧 Get started today: 💡 Read the full blog: #neocloud# #AI# #ModelStudio# #NVIDIAAI#

Latest open artifacts (#20#): New orgs! New types of models! With Nemotron Super, Sarvam, Cohere Transcribe, & others The top end of the market was quiet, but "industry-scale tinkering" just got very loud. We're seeing a massive shift: specialized, cheap open models are now the crucial tools complementing closed agents. Analysis & breakdowns of: - @NVIDIAAI Nemotron-3-Super-120B: 1M context + NVFP4. - @cohere Transcribe - @SarvamAI -105B - @MistralAI -Small-4-119B Massive coverage of new artifacts from: - Coding & Logic: @zeddotdev @Meituan_LongCat & Goedel-LM - Multimodal/OCR: @Microsoft @YuanAI_Lab @BaiduAI & @xiaohongshu (RedNote) - RAG & Search: @trychroma @LightOnIO & @miromind_ai - Robotics/Agents: @Reka_AI @hcompany_ai & @ServiceNowRSRCH - Creative & Audio: @bfl_ml @hume_ai & GAIR - Data & Infra: @allen_ai @IBM @markov__ai @StepFun_ai By @natolambert & @xeophon