MiniMax-H3-Image-PE-NVFP4-MTP-GGUF

MiniMax-H3-Image-PE is MiniMax's prompt-expansion model for its MiniMax-H3-Image image generator. It is fine-tuned from Qwen3.8-27B and released in FP4. It turns a short request into a detailed image description and recommends an aspect ratio, returned as JSON with expanded_caption and aspect_ratio, with the output constrained to a schema by XGrammar. It runs with thinking on by default (reasoning_effort xhigh, 4,096 max tokens including reasoning), and the reasoning is returned separately. The weights cover text-to-image expansion. For image editing, the release provides only a task prompt and a message-assembly example for a vision-language model of your choice, and the card says editing support is still under development. In editing mode the model describes what is preserved and changed and returns a target_aspect_ratio. It can be served with SGLang (0.5.23 or later) or run locally through the Modular Diffusers blocks, where the FP4 weights use about 19 GiB of GPU memory. The Diffusers path runs activations in bf16, so its outputs differ slightly from the SGLang deployment. The card gives no benchmarks. The license is the MiniMax-H3-Image Community License Agreement, not Apache 2.0, and some uses need separate written authorization from MiniMax. The Qwen3.8-27B base is Apache 2.0.

MTP — Speculative Decoding: Yes

Multi-Token Prediction (MTP) GGUF is a specialized GGUF model file format extension that integrates speculative decoding directly into the model weights to significantly accelerate local inference. Unlike traditional speculative decoding which requires a separate, smaller "draft" model, MTP GGUF files include additional output heads within the main model architecture that predict multiple future tokens in a single forward pass.

Model Files

File Name	Quant Type	File Size	File Link	Description
MiniMax-H3-Image-PE-NVFP4-MTP.gguf	NVFP4 (with MTP)	17.9 GB	Link	NVFP4 quantization with Multi-Token Prediction (MTP). Highly optimized for NVIDIA Blackwell hardware accelerators (e.g., RTX Pro 6000, RTX 50-series, B200, B300), offering native hardware-accelerated FP4 Tensor Core speeds and faster generation.
MiniMax-H3-Image-PE.mmproj.gguf	mmproj	931 MB	Link	Multimodal projection file. Used for vision/image capabilities.

---

Without MTP vs. With MTP Decoding

Benchmark	Without MTP	With MTP
Model	MiniMax-H3-Image-PE-NVFP4-GGUF	MiniMax-H3-Image-PE-NVFP4-MTP-GGUF ✱
File Size	17 GB	17.9 GB
Average Generation Speed	74.94 tokens/s	112.64 tokens/s
Speculative Decoding	Disabled	MTP (draft_n_max=2)
MTP Tensors	Not Preserved	Preserved
Test Runs	5	5

Performance Improvement

Metric	Result
Speed Improvement	+37.70 tokens/s
Relative Performance Gain	+50.3%
Generation Speed Multiplier	1.50×

MTP speculative decoding delivers approximately 50.3% faster generation, increasing throughput from 74.94 to 112.64 tokens/s across five benchmark runs.

---

NVFP4 Test Run

Tested MiniMax-H3-Image-PE-NVFP4-MTP.gguf on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU with 96 GB of VRAM, using llama-cli with full GPU offloading and MTP speculative decoding.

Performance Metrics

Metric	Value
Average Prompt Processing Speed	307.96 tokens/s
Average Generation Speed	112.64 tokens/s
Total Wall Time	66.0 seconds
Context Length	8,192 tokens
KV Cache (K)	q8_0
GPU	NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU VRAM Capacity	96 GB
Peak GPU VRAM Usage	16.34 GiB
Peak GPU Utilization	95%
GPU Layers Offloaded	99 (-ngl 99)
Random Seed	42
Speculative Decoding	MTP (draft_n_max=2)
Test Runs	5

Test Shots

Shot	Test
Shot 1	Image Edit: Day to Cyberpunk Night
Shot 2	Image Edit: Studio Product Relighting
Shot 3	Image Edit: Summer to Winter Wonderland
Shot 4	Image Edit: Vintage 1970s Film Portrait
Shot 5	Image Edit: Watercolor Painting Restyle

Run Inference

./build/bin/llama-cli \
  -m /path/to/models/MiniMax-H3-Image-PE-NVFP4-MTP.gguf \
  --spec-type draft-mtp \
  --spec-draft-n-max 2 \
  -ngl 99 \
  -c 8192 \
  --cache-type-k q8_0 \
  --seed 42 \
  -n 8192 \
  -st \
  -p "Your prompt here"

Performance results reflect this specific hardware and inference configuration. Actual throughput may vary depending on the prompt, model, software build, and runtime settings.

Snapshots (5)

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

MTP-repackaged and quantized using llama.cpp by — prithivMLmods.