GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
cbert33
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
cbert33
Model Tree
XiaomiMiMo/MiMo-V2.6-Flash-MOPD
Input Modalities
Output Modalities
Supported Functionality
attn.o_proj across all 48 transformer
layers. No other module class was selected for ablation.model.mtp.* tensors. The separate
upstream model_mtp.safetensors is therefore intentionally not duplicated.The transformation used Heretic v0.1.1
with 100 mlabonne/harmless_alpaca prompts and 100
mlabonne/harmful_behaviors prompts. Keyword-rate evaluation used the 100
harmful prompts; KL-divergence evaluation used 20 harmless prompts.
| Setting | Value |
|---|---|
| Trials | 1 |
| Target modules | 48 × attn.o_proj |
direction_index | 38.49 |
attn.o_proj.max_weight | 0.86 |
attn.o_proj.max_weight_position | 46.53 |
attn.o_proj.min_weight | 0.10 |
The selected trial produced KL divergence 0.0011. The keyword metric was
0/100 before and after the pass, so it was saturated at baseline and should
not be read as a behavioral or safety score. The useful measured result is the
small first-token distribution shift while limiting the edit to the attention
output projections.
| Setting | Value |
|---|---|
| EXL3 version | 1.5.3 |
| Average target bitrate | 4.0 bpw |
| Codebook | MCG |
| LM-head bitrate | 8 bits |
| MTP bitrate | 4 bits |
| Vision bitrate | 16 bits |
| Calibration matrix | 250 rows × 2,048 columns |
| Hessian regularization | 0.025 |
| Output scales | automatic |
The rank-sliced EXL3 payload contains 290,356 tensors and 158,980,524,144 bytes across 24 target shards. FP8 KV calibration adds 96 scale tensors: one K and one V scale for each of 48 layers. Those scales were derived from a 17,468-token calibration capture with eight observations per layer per rank, using a 10% headroom factor. The final index contains 290,452 tensors.
conversion-args.json, quantization_config.json,
rank-slice-manifest.json, KV_CACHE_CALIBRATION.json, and SHA256SUMS
record the conversion, slicing, calibration, and file-integrity details.
The included Xiaomi drafter is served with eight proposals per verification step. In a 305-step validation run against this exact Heretic EXL3 target, the first four proposal positions were accepted at 71.1%, 40.0%, 22.0%, and 11.1%. Overall acceptance across all eight positions was 20.5% (491 of 2,396 draft tokens). Text generation, structured tool calls, continuation after tool results, and long-prefix cache reuse also passed. These are validation observations from one two-node DGX Spark deployment, not general benchmark claims; broader performance testing remains in progress.
Stock vLLM and Transformers do not understand this rank-sliced tensor
schema. To run this we've created a custom version of vLLM using all native
sm120/sm121 libraries. Use: cbert33/cb-vllm-dual-dgx.
It is a vLLM 0.30-based two-node DGX Spark runner with rank-sliced EXL3/Trellis loading, DFlash, no async scheduling, FP8 KV cache, prefix caching, reasoning, and structured tool-call support. Both ranks must use the same runner revision, image, checkpoint revision, drafter, and runtime flags. The runner repository contains the build and deployment contract; this repository contains the model artifacts.
dflash/: Xiaomi's matching DFlash configuration, implementation, mask
embedding, index, and draft weights.audio_tokenizer/: Xiaomi's matching audio tokenizer configuration,
template, generation settings, tokenizer settings, and weights.assets/: the architecture and tool-call-repetition figures referenced by
the upstream model card.[!IMPORTANT] This is the MOPD upgrade of the MiMo-V2.6-Flash-RL checkpoint.
- MOPD2 (👉 Technical Report §5.6) Fuses several domain-specialized teachers into one model, extending to domains where reliable training-time verification is hard, such as long-horizon game development, scientific research and embodied intelligence.
- Diagnosing and Mitigating Tool-Call Repetition in MiMo-V2.6 (👉 Technical Blog) An easy-to-overlook failure mode in which the model keeps issuing the same or highly similar tool calls, appearing busy while making no progress. Nothing fails outright, so it tends to go unnoticed. The MOPD stage handles it efficiently, with a short specialized-teacher run that converges quickly.
MOPD2 distills several domain-specialized teachers into the student on-policy. The teachers fall into two families: mixRL teachers, trained on verifiable tasks, and SFT teachers, trained on synthetic demonstrations for open-domain tasks where a reliable reward is hard to design. Three streams contribute to a single update:
Method details are in Technical Report §5.6.
Following the release of MiMo-V2.6, tool-call repetition emerged as one of the most noticeable issues in agentic settings: the model would sometimes issue the same or highly similar tool calls repeatedly, consuming time and context without making progress. This checkpoint mitigates it.
Figure: response-level repetition rate on MiMo-V2.6-Flash, RL-stage versus this checkpoint, across context lengths and agent harnesses.
The technical blog has the full diagnosis. The fix is lightweight to train: a short specialized-teacher run that folds into the normal MOPD pass.
Figure 1. MiMo-V2.6 architecture.
| Model | Download |
|---|---|
| MiMo-V2.6-Pro-RL | 🤗 HuggingFace · 🤖 ModelScope |
| MiMo-V2.6-Flash-RL | 🤗 HuggingFace · 🤖 ModelScope |
| MiMo-V2.6-Pro-MOPD | 🤗 HuggingFace · 🤖 ModelScope |
| Component | MiMo-V2.6-Flash-MOPD |
|---|---|
| Layers (Total / SWA / GA) | 48 / 39 / 9 |
| Hidden Size | 4096 |
| SWA Heads (Q/KV) | 64 / 8 |
| GA Heads (Q/KV) | 64 / 4 |
| Head Dimensions (QK / V) | 192 / 128 |
| Sliding Window Size | 128 |
| Routed Experts (Total / Activated) | 256 / 8 |
| Max Context Length | 1M |
| MTP / Speculative Decoder |
The first Transformer block uses global attention with a dense FFN. Remaining blocks interleave local SWA and GA; both use sparse MoE FFNs without shared experts.
| Configuration | Value |
|---|---|
| Layers (Total / SWA / GA) | 28 / 24 / 4 |
| Hidden Size | 1280 |
| Attention Heads (Q / KV) | 32 / 8 |
| Head Dimension | 64 |
| Patch Size (T × H × W) | 2 × 16 × 16 |
| Sliding Window (Left / Right) | 64 / 64 |
| Spatial Merge Size | 2 × 2 |
| Parameters | 681M |
AudioTokenizer encoder: 24 layers (12 SWA / 12 GA), hidden 1024, 20 RVQ codebooks, 308M parameters. Audio patch encoder: 6 layers, 127M parameters; four frames per patch (25 Hz → 6.25 Hz).
5-layer SWA MTP drafter (DFlash-style). Predicts 7 subsequent tokens per forward pass for parallel verification.
For best performance, follow the SGLang MiMo cookbook. Docker image: lmsysorg/sglang:latest.
bash
sglang serve \ --trust-remote-code \ --model-path XiaomiMiMo/MiMo-V2.6-Flash-MOPD \ --tp 8 \ --dp 2 \ --enable-dp-attention \ --enable-dp-lm-head \ --mm-enable-dp-encoder \ --mem-fraction-static 0.65 \ --chunked-prefill-size 16384 \ --speculative-algorithm EAGLE \ --speculative-num-steps 3 \ --speculative-eagle-topk 1 \ --speculative-num-draft-tokens 4 \ --enable-multi-layer-eagle \ --reasoning-parser mimo \ --tool-call-parser mimo \ --host 0.0.0.0 \ --port 30000Follow the vLLM MiMo-V2.5 recipe. Stable vLLM may lag; pre-built image: docker pull vllm/vllm-openai:mimov25-cu129.
bash
vllm serve XiaomiMiMo/MiMo-V2.6-Flash-MOPD \ --tensor-parallel-size 4 \ --trust-remote-code \ --gpu-memory-utilization 0.95 \ --max-model-len auto \ --reasoning-parser mimo \ --tool-call-parser mimo \ --enable-auto-tool-choice \ --generation-config vllmRecommended sampling: temperature=1.0, top_p=0.95.
Also available in AI Studio, MiMo Code, Xiaomi MiMo Desktop, Xiaomi MiMo Open Platform API, and OpenRouter.
bibtex
@misc{mimo2026v26flashmopd, title={MiMo-V2.6-Flash-MOPD}, author={{Xiaomi MiMo Team}}, year={2026}, howpublished={\url{https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-MOPD}},}For questions or feedback, reach us at mimo@xiaomi.com or join our community:
attn.o_proj.min_weight_distance | 1.01 |
| Tensor-parallel slices | 2 |
| MiMo-V2.6-Flash-MOPD | 🤗 HuggingFace · 🤖 ModelScope |
| 5 SWA layers, window 1024 |