MiniMax-M2.7-speedy-colibri-nvfp4

MiniMax-M2.7-speedy-colibri-nvfp4

NVFP4 container for colibrì, a Rust MoE inference engine for single-box streaming-expert inference.

This repository exists so the model can be downloaded and run directly, with no conversion step. Converting the upstream checkpoint requires holding the source and the container on disk at the same time; this container is the finished artefact.

• Upstream: nvidia/MiniMax-M2.7-NVFP4
• Base model: MiniMaxAI/MiniMax-M2.7
• Architecture: 62 layers, GQA (48Q/8KV, partial rope 64); 256 experts top-8, SwiGLU
• Expert format: NVFP4

Use

The engine is github.com/GriffinPilz/SpeedyColibri. With it cloned and built, one command downloads this container and one serves it:

scripts/fetch.sh minimax-m2.7
scripts/serve.sh minimax-m2.7 8080          # OpenAI-compatible API

serve.sh fetches the container itself when it is missing, so the first line is optional — it is there for when you want the download to happen at a time you chose. Both are idempotent: over a complete directory they verify it, over an interrupted one they finish it. This one is 122 GB — budget the disk.

For a one-shot generation without the server, coli gen MiniMax-M2.7-container "<prompt tokens>".

Pulling it by hand works too, but the directory name is what the registry resolves, so use exactly this one (under your COLI_MODELS_ROOT, default ~/models) or serve.sh will not find it:

huggingface-cli download Kanposer/MiniMax-M2.7-speedy-colibri-nvfp4 --local-dir MiniMax-M2.7-container

Measured on one DGX Spark (GB10, 121.7 GiB), 2026-08-09

Single sequence, greedy, 512-token prompt. Median of repeated runs on one build, gated on token identity — a "faster" number that changed the output fails loudly instead of being reported as a win.

decode is the noisiest column here. The eight reps behind that median spanned 5.01-5.92 tok/s forward-only in one session — an 18% spread with no code change. Treat a single suite as one draw, and do not read a difference against an earlier revision of this card as a speedup.

Method, caveats and the full fleet table: SpeedyColibri README.

What was changed

Weights were repacked into colibrì's container layout. Routed experts are NVFP4 as published upstream; the container repacks them into coalesced per-expert spans for streaming. No fine-tuning, distillation, or other modification of model behaviour was performed.

Licence and attribution

This is a derivative of nvidia/MiniMax-M2.7-NVFP4 and is distributed under the upstream licence, nvidia-software-and-model-evaluation-license (licence text). The upstream publishes its licence as a URL rather than a file, so it is linked above rather than copied here.

All rights in the original model remain with its authors. Please read and comply with the upstream licence before downloading or using these weights.