Add MiniMax-M3 NVFP4 variant (MTP + non-MTP)

[Ankur-singh]

Adds an NVFP4 variant (nvidia/MiniMax-M3-NVFP4) to the MiniMax-M3 recipe — NVIDIA-quantized 4-bit weights, ~1/4 the BF16 VRAM (vram_minimum_gb: 257), fitting a single Blackwell (B200/B300) node with KV-cache headroom.

Both MTP and non-MTP are covered: the recipe's existing opt-in spec_decoding feature already wires the EAGLE3 draft head (Inferact/MiniMax-M3-EAGLE3, 3 speculative tokens), so the NVFP4 target runs with spec decoding on (MTP) or off (non-MTP) via the command builder — no new feature needed.

Changes to models/MiniMaxAI/MiniMax-M3.yaml:

• New nvfp4 variant (model_id: nvidia/MiniMax-M3-NVFP4), following the existing MiniMax-M2.7 NVFP4 precedent.
• Guide section "Quantized Variant (NVFP4, Blackwell)" with TP8 and +EAGLE3 (MTP) serve commands.
• meta description/date refresh; NVFP4 / EAGLE3 / vLLM PR links added to References.

NVFP4 support is in-flight in vLLM (PR #46380); the guide notes the build requirement until it merges.

Validated with node scripts/build-recipes-api.mjs (✓ JSON API parses; nvidia/MiniMax-M3-NVFP4 renders as a promoted variant, Blackwell-only).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jun 25, 2026 8:15pm

[gemini-code-assist]

Code Review

This pull request adds the new nvfp4 quantized variant (nvidia/MiniMax-M3-NVFP4) for the MiniMax-M3 model, updating the configuration and documentation to support running it on Blackwell GPUs with vLLM. The review feedback recommends defining and prefixing commands with the VLLM_USE_FLASHINFER_MOE_FP4=1 environment variable to ensure the model utilizes optimized FlashInfer kernels for FP4 MoE layers on Blackwell.

[Ankur-singh]

@faradawn can you please review this one?

[faradawn]

LGTM.

I've resolved the Gemini comments since the NVFP4 MOE env var is not needed. Below is the full serving command that supports this recipes

vllm serve nvidia/MiniMax-M3-NVFP4 \
$PARALLEL_ARGS \
--gpu-memory-utilization 0.90 \
--max-model-len $MAX_MODEL_LEN \ >>> not needed. 
--block-size 128 \ >>> not needed
--language-model-only \ >>> not needed
--max-cudagraph-capture-size 2048 \ >>> not needed
--max-num-batched-tokens "$((ISL * 2 ))" \ >>> not needed
--stream-interval 20 --no-enable-prefix-caching \ >>> not needed
--trust-remote-code > $SERVER_LOG 2>&1 &

[faradawn]

Suggested change

	## Quantized Variant (NVFP4, Blackwell)
	[`nvidia/MiniMax-M3-NVFP4`](https://huggingface.co/nvidia/MiniMax-M3-NVFP4) is
	an NVFP4 checkpoint quantized by NVIDIA — roughly **1/4 the VRAM** of the BF16
	release, so the 427B model fits comfortably on a single Blackwell node (B200 /
	B300) with KV-cache headroom. Select the **nvfp4** variant above, or pass the
	repo id directly to `vllm serve`.
	> **vLLM support is in-flight.** MiniMax-M3 NVFP4 needs the modelopt NVFP4 path
	> added in [vLLM PR #46380](https://github.com/vllm-project/vllm/pull/46380),
	> which is not yet merged. Until it lands in a release, build vLLM from that
	> branch (or a nightly once merged); a stock build will not recognise the NVFP4
	> quant config.
	```bash
	vllm serve nvidia/MiniMax-M3-NVFP4 \
	--tensor-parallel-size 8 \
	--block-size 128 \
	--tool-call-parser minimax_m3 \
	--reasoning-parser minimax_m3 \
	--enable-auto-tool-choice
	```
	Add `--enable-expert-parallel` (TP+EP) or `--data-parallel-size 8
	--enable-expert-parallel` (DP+EP) to scale across the node, exactly as for the
	BF16/MXFP8 commands above. For text-only serving, add `--language-model-only`
	to skip the vision encoder and free VRAM for KV cache.
	### NVFP4 + EAGLE3 spec decoding (MTP)
	The NVFP4 target pairs with the same EAGLE3 draft head as the other variants.
	Enable the **Spec decoding** feature above, or append the draft config to the
	command:
	```bash
	vllm serve nvidia/MiniMax-M3-NVFP4 \
	--tensor-parallel-size 8 \
	--block-size 128 \
	--tool-call-parser minimax_m3 \
	--reasoning-parser minimax_m3 \
	--enable-auto-tool-choice \
	--speculative-config '{"method": "eagle3", "model": "Inferact/MiniMax-M3-EAGLE3", "num_speculative_tokens": 3, "attention_backend": "FLASH_ATTN"}'
	```

[faradawn]

The information correctly matches the selection UI's command. Do we want to keep this? Otherwise, it looks good to me. @esmeetu

[esmeetu]

doesn't harm to have this.

[functionstackx]

following this recipe, it passes evals https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28197637552/job/83557638525