MiMo V2.6 Pro MOPD
Run this prepared Flare runtime on eight NVIDIA SM120 GPUs with 96 GB VRAM each. It supports 64 concurrent requests, seven MTP draft tokens per round, text and images, and a 1,048,576-token per-request context limit. Serving uses Flare's native Rust runtime without Python or runtime compilation.
The served model takes text, images, video clips and audio in one request. Its feature arena accepts 32,768 positions per request, so a 4K page costs about 8,160 image tokens and a 30-second 720p clip about 14,080 video tokens. The package contains model weights, compiled CUDA kernels and captured execution graphs. It requires the Flare image below; Transformers, vLLM, SGLang and llama.cpp cannot load this runtime format. Expert weights use NVFP4 and KV storage uses FP8 compression.
Use Linux x86-64, the NVIDIA Container Toolkit and working CUDA peer access between all eight GPUs. Eight RTX PRO 6000 Blackwell Workstation Edition GPUs were used for serving checks. Prefer GPUs behind the same PCIe switch. Allow about 537 GiB (577 GB) of disk space for the runtime files, plus download-cache headroom. The reference deployment keeps a 64 GiB host-side prefix cache; allow at least 96 GiB of host RAM for that cache and serving overhead.
The public main revision contains the model card, license and publisher profile. Runtime files are in gb202. Keep the published layout intact: model.safetensors is the package manifest, the payloads live under weights/tp8/rank-0..7/ and deployments/gb202-tp8/, and publisher.json stays beside them.
Download and run
Install the Hugging Face hf CLI and download the complete C64 runtime:
mkdir -p models flare-state
hf download mitomtuna/MiMo-V2.6-Pro-MOPD-FLARE \
--revision gb202 \
--local-dir ./models
The branch also carries the card files; the runtime only needs model.safetensors, weights/, deployments/ and publisher.json, all under the directory passed to run. This package needs the Flare image built from the 2026-10-27 engine β ghcr.io/tunamitom/mimo-pro-flare:latest, pinned by digest through the 74eabe8f tag.
Run the primary file in Docker:
docker run --pull=always --rm -it --user "$(id -u):$(id -g)" \
--gpus '"device=0,1,2,3,4,5,6,7"' --ulimit memlock=-1:-1 \
-p 127.0.0.1:8000:8000 \
--mount "type=bind,src=$PWD/models,dst=/models,readonly" \
--mount "type=bind,src=$PWD/flare-state,dst=/state" \
ghcr.io/tunamitom/mimo-pro-flare:latest \
run /models/model.safetensors \
--publisher-keys /models/publisher.json \
--gpus 0,1,2,3,4,5,6,7 --max-concurrency 64 --mtp-draft-tokens 7 \
--mtp-adaptive-verification --scheduler-policy throughput \
--kv-cache-limit memory --trust-store /state/trust.toml \
--host 0.0.0.0 --port 8000
Review and accept the publisher and model license when prompted. The writable state directory preserves that acceptance. This build is signed with the local-development publisher profile shipped as publisher.json. Docker selects host GPUs 0β7; Flare addresses them by their container-local indices.
Download through Flare
Flare can select gb202, download the complete package and serve it from a persistent cache. The explicit primary filename selects this C64 runtime:
mkdir -p flare-cache flare-state
docker run --pull=always --rm -it --user "$(id -u):$(id -g)" \
--gpus '"device=0,1,2,3,4,5,6,7"' --ulimit memlock=-1:-1 \
-p 127.0.0.1:8000:8000 \
--mount "type=bind,src=$PWD/flare-cache,dst=/cache" \
--mount "type=bind,src=$PWD/flare-state,dst=/state" \
ghcr.io/tunamitom/mimo-pro-flare:latest \
run mitomtuna/MiMo-V2.6-Pro-MOPD-FLARE \
--revision gb202 --filename model.safetensors \
--gpus 0,1,2,3,4,5,6,7 --max-concurrency 64 --mtp-draft-tokens 7 \
--mtp-adaptive-verification --scheduler-policy throughput \
--kv-cache-limit memory --cache-dir /cache \
--trust-store /state/trust.toml --host 0.0.0.0 --port 8000
For download only, replace run with download and omit the serving flags. This repository is public. If authentication is required, run hf auth login, mount the token file read-only and pass its container path with --token-file.
Send a request
The OpenAI-compatible API accepts model name MiMo-V2.6-Pro-MOPD:
curl --fail http://127.0.0.1:8000/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"MiMo-V2.6-Pro-MOPD","messages":[{"role":"user","content":"Explain how a hash table works."}],"max_tokens":1024}'
Sampling defaults are fixed at temperature 1.0, top-p 0.95, top-k disabled, min-p 0, repetition penalty 1.0, and zero presence/frequency penalties; the seed is request-controlled.
Images use OpenAI image_url content items with PNG, JPEG or static WebP data URLs. Image tokens share the model's 1,048,576-token context, and each image is encoded in a 16,384-token feature window. Text prompts use the model's Jinja chat template, which also handles tool calls and think-tag reasoning.
Serving settings
--max-concurrencydefaults to 64. Keep 64 for aggregate throughput; use a lower value to reduce resident execution state. The package cannot exceed 64 active requests.--mtp-draft-tokensdefaults to 7, with adaptive verification enabled. Keep these settings for the packed graph catalog. The catalog supports widths 1β8; target-only execution uses--mtp-draft-tokens 0.--scheduler-policyβ the examples use throughput for concurrent generation;balancedfavours per-request latency.--kv-cache-limit memorykeeps the KV pool in device memory; the 64 GiB prefix cache in the example is host-side and optional.
Model tree for mitomtuna/MiMo-V2.6-Pro-MOPD-FLARE
Base model
XiaomiMiMo/MiMo-V2.6-Pro-MOPD