just dropped
https://huggingface.co/nvidia/Qwen3.6-27B-NVFP4
9:57 AM · Jun 30, 2026 · 149 Views
Nvidia releases Qwen3.6-27B-NVFP4, using its Blackwell-era 4-bit format to reduce memory requirements by 4x
The release supports high-throughput inference via vLLM integration
439057.6K
Sentiment
Positive users express excitement about trying NVIDIA's new FP4-quantized Qwen model on their RTX hardware while negative users note disappointment over missing features like tcgen05 on sm120.
Pos
50.0%
Neg
50.0%
3 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Related links
Posts from X
Most Activity
Most Activity
VIEWS3.4KBOOKMARKS4REPLIES3
Brian Roemmele@BrianRoemmele
Boom!
Nvidia Drops Qwen3.6-27B-NVFP4 on Hugging Face
Hugging Face CTO Julien Chaumond announced today that Nvidia has released the Qwen3.6-27B-NVFP4 model, a highly efficient quantized version of Alibaba’s recently launched 27-billion-parameter Qwen3.6 series.
NVFP4 is Nvidia’s proprietary 4-bit floating-point format introduced with the Blackwell GPU architecture.
It delivers up to 4× memory reduction compared to FP16 while preserving strong accuracy, making large models far more practical on Blackwell-based systems like DGX Spark and RTX PRO 6000 cards.
New benchmarks show excellent throughput with tools such as vLLM, especially when combined with speculative decoding techniques like MTP.
The release gives developers immediate access to a powerful, coding-focused 27B model that runs efficiently on consumer and enterprise Blackwell GPUs.
It joins a growing ecosystem of Qwen3.6 quantizations and underscores Nvidia’s push to accelerate open AI model deployment on its platform.
Julien Chaumond@julien_c
just dropped
https://huggingface.co/nvidia/Qwen3.6-27B-NVFP4
4hViews 3.4KLikes 15Bookmarks 4
LIKES16
Julien Chaumond@julien_c
Remember that recent llama.cpp builds can dispatch GEMMs directly to the Blackwell FP4 Tensor Cores.
Lotto@LottoLabs
Ohhh we need some Blackwell bros to run this!
Nvidia coming through with the goat quants
https://huggingface.co/nvidia/Qwen3.6-27B-NVFP4
3hViews 2.8KLikes 16Bookmarks 1
RETWEETS1
Noé Flandre@NoeFlandre
@julien_c In case you don’t know, as I didn’t, NVFP4 is Nvidia’s floating point format. It was introduced with the Blackwell GPU architecture. It reduces memory by up to 4x vs FP16 while preserving accuracy
8hViews 805Likes 3
Sebastian S. Cocioba🪄🌷@ATinyGreenCell
@julien_c Cant wait to try this on my http://tiiny.ai thingggg
4hViews 330Bookmarks 2
Raghuvaran_mr@RaghuModupalli
@julien_c Will it work for AMD ?? I guess no ??
7hViews 842Likes 2
Julien Chaumond@julien_c
❤️❤️
matheus.@matheusbuild
peut on imaginer une vie sans hugging face 🤗 Il n’y a plus de retour en arrière
3hViews 1.2KLikes 7Bookmarks 0
TrueStory@quantumleap68
@julien_c any dgx spark TPS numbers on it?
7hViews 455Likes 2
P A@PA80339646
@julien_c Is there a NVFP8 blessed by Nvidia?
6hViews 672Likes 1
Adam 🤗@AdamMolnarHF
@julien_c big if true
6hViews 362
CBir@c__bir
@julien_c works with mtp as well?
7hViews 346
Jakove_HR@JakoveHr
@RaghuModupalli @julien_c Not nvfp4 bro definitly Not its Even Blackwell only
7hViews 51
Jakove_HR@JakoveHr
@RaghuModupalli @julien_c Get gguf i guess llamacpp
7hViews 50
42loops@42loops
@NoeFlandre @julien_c Je crois que ma rtx 6000 pro va être super contente…
et si je capte bien je peux maintenant la remplacer par seulement une rtx 3090 ?
Si c’est ça c’est une putain de bonne nouvelle
7hViews 37
Zack Angelo@zackangelo
@julien_c no tcgen05 on sm120 though :(
2hViews 33
jon3k@jon3k
@RaghuModupalli @julien_c Just run an unsloth gguf at whatever quant you want.
3hViews 15
Products Heart@productsheartco
@julien_c !!
2h
QDA HUB@QDAHUB
@julien_c We’ve Launched! [QDA-HUB] is here to transform how businesses track and scale. Our flagship Shopify Performance Dashboard delivers premium analytics through our know made dashboard — giving companies clarity, speed, and measurable impact. link:- https://qda-hub.vercel.app/
7h