No official torrent yet — requesting it tells the flock what to host next. You cannot seed a catalog-only page.
Cataloged from XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash
All models by XiaomiMiMo →
🎨 Xiaomi MiMo API Platform (Request Access) | 🗨️ Xiaomi MiMo Studio (Free Trial)
MiMo-V2.5-Pro-FP4-DFlash is the underlying model that powers MiMo-V2.5-Pro-UltraSpeed:
Together they cut both the per-parameter bit width and the number of backbone forward passes, the two dominant costs of trillion-parameter decoding.
At the trillion-parameter (1T) scale, even 8-bit (FP8/INT8) inference carries severe memory-footprint and memory-bandwidth costs. Lowering the parameter bit width translates directly into faster decoding. We therefore adopt FP4 quantization and block-diffusion speculative decoding. Key features of this release:
We quantize only the MoE experts to MXFP4 (block size 32) and keep attention projections and other modules at higher precision (the attention o_proj of every layer is excluded from FP4). With FP4 QAT, quality stays close to the FP8 baseline:
Conventional speculative decoding relies on a small draft model to guess the next tokens, which the large model then verifies; the rejection-sampling verification keeps the output lossless. Its bottleneck is that draft quality bounds the acceptance rate, while a stronger draft costs more compute.
To break this trade-off we adopt the block-level masked parallel-prediction approach DFlash: the draft fills an entire block of masked positions in one forward pass. We landed this on MiMo-V2.5-Pro with custom optimizations for trillion-scale MoE and long-context serving, using the Muon second-order optimizer and model self-distillation so that even a small mask block keeps a strong acceptance rate while pushing the draft-stage cost close to its limit:
In practice, we further cap the mask block size at 8 to lower verification overhead and raise concurrency.
DFlash inference with the FP4 backbone is supported in SGLang. The drafter is launched alongside the backbone via the speculative-decoding flags and inherits the backbone's tensor/expert-parallel topology.
The following is an example of running the model with SGLang. Point --model at this repository and --speculative-draft-model-path at its dflash/ subdirectory.
For questions or feedback, reach us at mimo@xiaomi.com or join our community:
python3 -m sglang.launch_server \ --model MiMo-V2.5-Pro-FP4-DFlash \ --speculative-algorithm DFLASH \ --speculative-draft-model-path MiMo-V2.5-Pro-FP4-DFlash/dflash \ --speculative-num-draft-tokens 8 \ --ep-size 16 \ --tensor-parallel-size 16 \ --data-parallel-size 2 \ --enable-dp-attention \ --enable-dp-lm-head \ --quantization fp8 \ --attention-backend fa3 \ --moe-dense-tp-size 1 \ --dtype bfloat16 \ --mem-fraction-static 0.65 \ --context-length 65536 \ --page-size 1 \ --trust-remote-code \ --disable-overlap-schedule \ --skip-server-warmup \ --dist-init-addr ${MASTER_ADDR}:20000 \ --nnodes ${WORLD_SIZE} \ --node-rank ${RANK} \ --host 0.0.0.0 \ --port 29999@misc{mimo2026v25pro_fp4dflash, title={MiMo-V2.5-Pro-FP4-DFlash}, author={{Xiaomi MiMo Team}}, year={2026}, howpublished={\url{https://huggingface.co/collections/XiaomiMiMo/mimo-v25}},}