license: mit base_model: dealignai/MiMo-V2.6-Flash-RL-UNCENSORED pipeline_tag: text-generation library_name: gguf quantized_by: cafonez language:

• en
• zh tags:
• gguf
• mimo
• mimo2
• mimo-v2-6-flash
• uncensored
• mxfp4
• moe
• rocmfpx
• vulkan
• strix-halo
• gorgon-halo
• framework-desktop

MiMo-V2.6-Flash-RL UNCENSORED GGUF for the Framework Desktop (Gorgon Halo)

A 4.29 bpw GGUF of dealignai/MiMo-V2.6-Flash-RL-UNCENSORED, the refusal-removed edition of XiaomiMiMo/MiMo-V2.6-Flash-RL (309.8B total parameters, MoE with 256 experts, top-8). It is sized to run fully on the GPU of the 192 GB AMD Ryzen AI Max 400 ("Gorgon Halo") Framework Desktop, and was built and tested on that machine.

• Requires ROCmFPX main (from PR #33 on). The routed experts are stored as one fused ffn_gate_up_exps tensor per layer, which stock llama.cpp, LM Studio and Ollama do not load for MiMo yet.
• Size: 154.6 GiB (166.0 GB) in 4 shards. Pass the first shard to -m; the rest are found automatically.
• Text only: the vision and audio encoders of the original are not included.
• Uncensored: refusals were removed by dealignai at the weight level (see their card). You are responsible for how you use it.

Files

Quantization recipe

• The routed experts are the checkpoint's own MXFP4 weights, repacked into GGUF MXFP4 without requantization.
• Attention and the dense/MTP projections were quantized to Q5_K with an importance matrix computed on a mixed code/prose calibration set. This cut the bytes read per generated token and raised decode speed from about 20 to 23.8 tok/s.
• Each layer's gate and up experts are stored as one tensor (a byte copy), which lets the GPU run one expert matmul instead of two.
• The three MTP (next-token prediction) layers are included.

Quality

KL divergence of this file against the same model with attention at Q8_0 (experts identical in both): mean KLD 0.0185, same top-1 token 94.5% (llama-perplexity --kl-divergence, Vulkan). Fusing gate/up and storing the router in F16 left the KLD unchanged.

Speed on the Framework Desktop (Gorgon Halo)

AMD Ryzen AI Max+ PRO 495 / Radeon 8065S, 192 GB LPDDR5X, Ubuntu 26.04, kernel 7.0, Mesa 26.0.8 RADV, Vulkan backend, ROCmFPX main. llama-bench -fa 1 -b 2048 -ub 2048, 3 runs:

These are local measurements on one machine, not leaderboard results.

How to run

1. Let the GPU use the memory

The weights need about 155 GiB of GPU-visible memory plus the KV cache. On a 192 GB machine, raise the GTT limit with kernel parameters (this gives the GPU 176 GiB) and reboot:

ttm.pages_limit=46137344 ttm.page_pool_size=46137344

On Ubuntu, add them to GRUB_CMDLINE_LINUX_DEFAULT in /etc/default/grub, run sudo update-grub, and reboot. Check with:

cat /sys/class/drm/card*/device/mem_info_gtt_total   # about 189000000000 bytes

Keep the BIOS GPU memory carve-out at its minimum: the GPU uses GTT, and memory carved out as VRAM is no longer available to it as GTT. Only run one model this size at a time; two will not fit in 192 GB.

2. Build ROCmFPX with Vulkan

sudo apt install build-essential cmake libvulkan-dev glslc mesa-vulkan-drivers
git clone https://github.com/ROCmFPX/ROCmFPX && cd ROCmFPX
cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j --target llama-server llama-cli

3. Download

hf download cafonez/MiMo-V2.6-Flash-RL-UNCENSORED-Gorgon-GGUF --local-dir MiMo-Gorgon-GGUF

4. Serve

./build/bin/llama-server \
  -m MiMo-Gorgon-GGUF/MiMo-V2.6-Flash-RL-UNCENSORED-MXFP4-Q5K-00001-of-00004.gguf \
  -dev Vulkan0 -ngl 999 -fa on \
  -c 131072 -b 2048 -ub 2048 -np 1 \
  --jinja --temp 0.7 --reasoning-budget 4096 \
  --host 127.0.0.1 --port 8080

• --jinja is required for the chat template (thinking and tool calls). Send "chat_template_kwargs": {"enable_thinking": false} in a request to turn thinking off.
• --temp 0.7: at the card default of 1.0 the model sometimes emitted hundreds of tool calls in one reply in agent use.
• --reasoning-budget 4096 caps thinking per turn; without it, agent turns can spend 15K+ characters thinking.
• For coding agents, also send "parallel_tool_calls": false with each request so the model makes one tool call per reply.
• -c 131072 fits comfortably with 176 GiB of GTT; -ub 2048 gives the best prefill.

Then use any OpenAI-compatible client against http://127.0.0.1:8080/v1.

Credits

• Model: Xiaomi MiMo-V2.6-Flash-RL (MIT).
• Uncensored edition: dealignai (MIT).
• MiMo support in llama.cpp and its contributors; fused gate/up loading in ROCmFPX PR #33.