(Corrected we are not sure about the model size) Hi everyone, Architectural Foundations & Design MiniMax M3 is a native multimodal model pre-trained on over 100 trillion tokens of interleaved multimedia data, allowing it to excel at complex tasks like agentic coding (scoring 59.0% on SWE-bench Pro). Structurally, it combines a highly sparse Mixture-of-Experts (MoE) backbone with a unique mechanism called MiniMax Sparse Attention (MSA) to efficiently handle a 1-million token context window. Instead of standard dense attention or lossy KV compression methods (like DeepSeek’s MLA), MSA preserves raw, uncompressed Key-Value representations to prevent semantic degradation. It partitions the KV-cache into fixed blocks and uses a lightweight Top-K router to isolate only high-relevance blocks per query, cutting compute costs to 1/20th. At the hardware level, it implements a smart KV-Block-Major memory layout to optimize SRAM locality. By loading a specific KV block once and processing all matching query clusters sequentially, it bypasses traditional HBM bandwidth bottlenecks, yielding a 9.7x prefill speedup and up to a 15.6x boost in decoding generation speeds. I hope vllm and atlas will add support soon i am curious on the first benchmarks. Best Regards

I wonder if nvidia planned for it. Otherwise a double nic is a waste. But makes sense for a ring topology

I am running this recipe on 3 nodes RIGHT NOW. It “works” but there is a lot of issues. I curently got it at 375K context. The main issue right now is it keeps getting stuck in loops when I ask a question the reasoning just LOOPS. 10tok/s I guess I can take a swing at the TP=3

How are you running it? Via RAY? But VLLM does not take tp=3? SGLang? Jeeez.. I can’t make a financial responsible reasoning to buy 2 more sparks + router + cable as returns are negligible for me, but 1 more + 1 basic cable for extra GPU and 128 GB? Hell yes if it works..

Unlikely, however one thing I will do over the next few weeks is to see if my theory of a fully meshed 4 node cluster is physically possible with this hardware. Should have the required cable for that proof of concept at some stage in the next two weeks (3-5 wees lead time sucks).

SGLang, PP=3 (pipeline parallel) — not Ray, not TP. You’re right that vLLM won’t do TP=3 on M3: the head dims don’t divide by 3 (64 attn / 4 KV heads), so tensor-parallel across 3 GPUs isn’t possible out of the box. That’s why I’m on sglang’s pipeline-parallel for now. NCCL runs over plain TCP between the 3 nodes — RoCE wouldn’t init for sglang’s PP coordinator on these CX7 NICs (the 2nd communicator crashes ~90s in), so I dropped back to TCP. That TCP pipeline is the ~10 t/s ceiling: PP means the nodes take TURNS on a single request instead of splitting it, so single-stream stays slow. Concurrency is where it shines (~90-100 t/s aggregate at 16 parallel). Running 375K context at MEMFRAC 0.90 (MSA KV can’t fp8-compress, so bf16 KV is the limit). The loops + the slow single-stream are exactly why I’m about to take a swing at Luke Alonso’s TP=3 commit (fb63c9aa). His “virtual sharding” pads the heads (64→96, KV 4→6) so they divide by 3 cleanly — which finally makes real TP=3 possible on a 3-Spark box, no 4th node. Downloading his NVFP4 quant now. The open question is whether the NVFP4 kernels run on GB10/SM121 (his card targets SM120). If they do, it should beat the PP=3 latency. Will report numbers.

Please keep us informed. This is incredible opportunity to get the hardware that can be used before price becomes astronomical.

For tp=3 you need to run the b12x developer’s VLLM fork which integrates it. I’ve got some more cables coming so will try it as well next week. It’s one of the release branches in here: GitHub - local-inference-lab/vllm: A high-throughput and memory-efficient inference and serving engine for LLMs · GitHub

[image] p33zy: For tp=3 you need to run the b12x developer’s VLLM fork which integrates it. I’ve got some more cables coming so will try it as well next week. It’s one of the release branches in here: GitHub - local-inference-lab/vllm: A high-throughput and memory-efficient inference and serving engine for LLMs · GitHub Thanks! Heads up though, I checked all 43 branches on local-inference-lab and none actually carry the M3 model code yet (just the b12x kernels + TP head-padding separately). Luke’s commit fb63c9a in vllm-project consolidates it all — full M3 model + TP=3 virtual sharding + b12x sparse-attn in one place. Building from that now, will report numbers.

Any idea if SGlang works well enough with GGUF to use the unsloth GGUF’s of M3?

tried IQ4_XS - result = 10.7 tok/sec

MiniMax M3 : NVFP4 for Quad DGX Spark

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

tonyd615 June 15, 2026, 1:54pm 117

Results still working Working recipe: MiniMax-M3 NVFP4 at TP=3 on 3x DGX Spark (no 4th node) + the OOM fixes

Topic		Replies	Views
MiniMax M3 NVFP4 and NVFP4 REAP 50 for 4x & 2x DGX Sparks DGX Spark / GB10 Projects	53	5243	July 2, 2026
MiniMax-M3 (428B MoE + vision) at ~14–15 tok/s on 2× DGX Spark — EAGLE3 speculative decoding is the unlock DGX Spark / GB10 Projects llama , deepseek	2	1264	July 4, 2026
Minimax3 on 2 nodes decode ~10.7 tok/s, 4bits DGX Spark / GB10 llama	26	2132	June 21, 2026
MiniMax M2.5 released (not available on HuggingFace as of now) -- is DGX Spark ready? DGX Spark / GB10	92	7814	April 12, 2026
MiniMax M2.7 NFVP4 Recipe & Benchmarks DGX Spark / GB10 llama	126	15222	August 20, 2026
MiniMax-M3-W4A16-GPTQ 2xGB10 Deployment (36 t/s) (fp8, nvfp4, KVarN, EAGLE-3) DGX Spark / GB10 agentic-ai	25	3561	August 17, 2026
MiniMax-M3-AWQ on 4× GB10, fp8 KV, 262k context, adaptive reasoning, ~30 tok/s DGX Spark / GB10 llama	17	1777	July 21, 2026
Get Running: nvidia/MiniMax-M3-NVFP4 - vLLM - No Ray - TP4 DGX Spark / GB10 Projects	36	1677	September 10, 2026
DGX Spark performance DGX Spark / GB10	49	8665	February 13, 2026
MiniMax 2.5 REAP - NVFP4 on single DGX Spark DGX Spark / GB10	25	3847	April 1, 2026

MiniMax M3 : NVFP4 for Quad DGX Spark

Related topics