Hi everyone, I have not tried this yet, but I want to share this with you. The first TurboQuant quantised LLM models appear on Hugging Face :-). [image] fno2010/MiniMax-M2.7-TQ3 · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source an…

Yeah, looks like one of the PRs (for KV cache is already merged upstream), and I can add the second one too. It is Python-only, so technically it could be added as a mod instead of baking it into the core build.

It will be great, if this works! Let me know when we can test it. Not urgent. Many many thanks!

Hi good sir, may I ask if you’ve made this work with a mod yet?

Not yet, have been catching up with other things :)

Hi @eugr , @eugr_nv , I wonder, do you think this can be added and when? This will be really useful for the community here. Many thanks in advance! CM

I believe, TurboQuant support has been merged into vLLM already, so it should be supported in the community docker. I haven’t tested it personally though.

Hi @eugr_nv , Partially. From GitHub - varjoranta/turboquant-vllm: TurboQuant+ KV cache compression for vLLM. 3.8x smaller KV cache, same conversation quality. Fused CUDA kernels with automatic PyTorch fallback. · GitHub . For KV cache yes. KV cache compression for GQA/MHA models (Qwen, Llama, Mi…

MiniMax M2.7 TQ3 - A TurboQuant 3-bit quantized version of MiniMax-M2.7 for single DGX Spark

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10

chrm May 29, 2026, 8:50pm 6

Hi @eugr , @eugr_nv ,
I wonder, is there any chance to implement this easily in the community docker image?
This would be great.

This might also benefit Step-3.7-Flash, because this model could be run on one DGX Spark.

Many thanks!

Kind regards,

Topic		Replies	Views
Why Turboquant saves DGX twice DGX Spark / GB10	134	13702	May 31, 2026
Introducing PrismaQuant DGX Spark / GB10	166	8514	June 12, 2026
What's the best speed we can get with Qwen 3.6 27B without quantizing? DGX Spark / GB10	64	28153	July 6, 2026
DGX Spark GB10 / vLLM 0.19.1: TurboQuant KV cache integration results on Qwen3.5 and Nemotron, including gather-free Triton decode and CUDA WPH decode DGX Spark / GB10 Projects nemotron	5	1965	April 7, 2026
Qwen3.5-122B-A10B on single Spark: 15 → 21.5 tok/s with hybrid GPTQ-INT4 + FP8 dense layers (https://github.com/rmstxrx/vllm-hybrid-quant) DGX Spark / GB10 cuda	9	1169	March 20, 2026
Introducing PrismaScout -- PrismaQuant v2! DGX Spark / GB10	138	10975	July 21, 2026
Qwen3.5-122B-A10B NVFP4 Quantized for DGX Spark — 234GB → 75GB, Runs on 128GB DGX Spark / GB10 Projects	44	13671	April 9, 2026
Qwen3.8-27B on DGX Spark using vllm: NVFP4 vs FP8 performance DGX Spark / GB10 llm	8	5543	August 24, 2026
Introducing Spark Auto Round /w OpenCode Instruct dataset DGX Spark / GB10 cuda , spark , agentic-ai	110	5475	July 17, 2026
Benchmark Report: Qwen/Qwen3.6-27B vs nvidia/Qwen3.6-27B-NVFP4 DGX Spark / GB10 Projects	5	2938	July 4, 2026

MiniMax M2.7 TQ3 - A TurboQuant 3-bit quantized version of MiniMax-M2.7 for single DGX Spark

Related topics