Xiaomi MiMo Releases DFlash Acceleration Scheme

karminski3 · x · 2026-07-13

Xiaomi released **MiMo-V2.5-DFlash** on Hugging Face; the code is already out even though the README hasn't been updated yet. The focus isn't on "how strong the new model is," but rather its **speculative decoding implementation**: the draft model uses **block diffusion** to guess an entire block of tokens in a single forward pass, which are then verified by the large model all at once. The author breaks down the code, noting that this is a completely different route from traditional autoregressive drafting like EAGLE. Implementation details mentioned: - In coding contexts, acceptance length can reach 6+, resulting in significant speedups - Draft weights are extracted separately, taking up about 2.94G locally - It's not a standalone small model; lacking embed / lm_head, it still requires KV injection from the main model's hidden state - `target_layer_ids` are aggressively set to `[0, 11, 23, 35, 47]` - `block_size=8` is more conservative than in papers and Qwen implementations, potentially offering more stability but slightly lower parallelism Overall, this is a technical teardown focused on an **inference acceleration plugin/deployment scheme** rather than just a new model announcement.

Related event: Xiaomi Releases MiMo-V2.5-DFlash Weights and Code(2 posts)→

Original post →

More from Infra

Infra channel →