Unsloth AI releases NVFP4 quantized Qwen3.6 models, offering up to a 2.5x speedup inside 24GB of VRAM · Digg

/TECH

Unsloth AI releases NVFP4 quantized Qwen3.6 models, offering up to a 2.5x speedup inside 24GB of VRAM

C🤗

The optimized models reach 17,561 tokens per second on Blackwell GPUs.

Combined views

2 posts, first seen 3h ago

52 likes

Unsloth AI releases NVFP4 quantized Qwen3.6 models, offering up to a 2.5x speedup inside 24GB of VRAM

C🤗

Commentary on X

Highest ranked

mr-r0b0t@mr_r0b0tTECH

@danielhanchen Re: unsloth/Qwen3.6-35B-A3B-NVFP4-Fast Happily confirming this is FAST FAST on a single @NVIDIAAI DGX Spark (GB10) ♥️♥️♥️ https://x.com/mr_r0b0t/status/2075604862562173203/photo/1

Combined views

2 posts, first seen 3h ago

52 likes comments

4 comments

13 saves

10 reposts

Sentiment

Positive100%0%Negative

Positive Read

Users are excited about Unsloth's NVFP4 quants for Qwen3.6 models because they deliver impressive 2.5x GPU speedups with strong real-world throughput on hardware.

Based on 4 visible X reactions from 10 accounts; directional sample.

Digg Deeper

Ask a question below.

Published answers will appear here.

Commentary on X

Highest ranked

mr-r0b0t@mr_r0b0tTECH

@MichaelHutu @UnslothAI IKR! GPT-5.6 should have no trouble getting these rolling 😀

Sparsely Activated Primate@SparselyActiveTECH

@mr_r0b0t @UnslothAI Unsloth's 27B almost matching Nvidia's 35B-A3B on throughput is wild

Vasko@RoliumGensTECH

@mr_r0b0t @UnslothAI we are ALL testing this today!!

View all

Daniel Han@danielhanchenTECH

2.5x faster NVFP4 Qwen3.6 27B + 35B-A3B quants! We applied our Unsloth dynamic method to NVFP4 using W4A4 vs W4A16. Also 2 35B-A3B versions - NVFP4 (1.56x faster) and NVFP4-Fast (1.79x faster). Fast uses pure W4A4, whilst non fast uses a mixture to stay a bit more accurate. https://twitter.com/UnslothAI/status/2075566124687892597

The optimized models reach 17,561 tokens per second on Blackwell GPUs.

mr-r0b0t@mr_r0b0tTECH

@MichaelHutu @UnslothAI IKR! GPT-5.6 should have no trouble getting these rolling 😀

Sparsely Activated Primate@SparselyActiveTECH

@mr_r0b0t @UnslothAI Unsloth's 27B almost matching Nvidia's 35B-A3B on throughput is wild

Vasko@RoliumGensTECH

@mr_r0b0t @UnslothAI we are ALL testing this today!!

View all

Daniel Han@danielhanchenTECH

13 saves

10 reposts

Sentiment

Positive100%0%Negative

Positive Read

Users are excited about Unsloth's NVFP4 quants for Qwen3.6 models because they deliver impressive 2.5x GPU speedups with strong real-world throughput on hardware.

Based on 4 visible X reactions from 10 accounts; directional sample.

Digg Deeper

Ask a question below.

Published answers will appear here.