AI models 1 min read

MiMo-V2-Flash vs Step 3.5 Flash: 15B active of 309B against 11B active of 196B

Two Mixture-of-Experts models with sliding window attention and 256K-class context compared on size, serving and licence notes.

Size and attention design

Both READMEs describe Mixture-of-Experts language models for reasoning and agentic work with sliding window attention and a 256K-class context. MiMo-V2-Flash has 309B total and 15B active parameters, interleaves sliding window and global attention at 5:1 with a 128-token window, and states a 256k context window.

Step 3.5 Flash activates 11B of its 196B parameters per token, uses a 3:1 sliding window ratio and states a 256K context window. Both describe multi-token prediction: the MiMo README says its module triples output speed, and the Step README describes 3-way prediction and a throughput of 100 to 300 tok/s in typical usage.

Licences and serving

Both repositories carry an Apache License, Version 2.0 file. The Step 3.5 Flash README says so in its license section, while the MiMo-V2-Flash README ends with a use-restrictions notice that lists uses the reader agrees not to make, including military use, and says it does not amend the licence. The MiMo-V2-Flash model card on Hugging Face, read on 9 November 2026, declares the MIT licence, which differs from the repository's Apache file, so confirm which terms govern its weights.

MiMo recommends SGLang and its launch example sets a tensor-parallel size of 8. Step 3.5 Flash lists vLLM, SGLang, Hugging Face Transformers and llama.cpp and states it can run on high-end consumer hardware such as the Mac Studio M4 Max. Its known-issues section mentions repetitive reasoning or mixed-language outputs in specialized domains and long dialogues.

Who should look at which

Readers who want the larger model and a Xiaomi-published base checkpoint alongside it may start with MiMo-V2-Flash, after reading its use-restrictions notice.

Readers who want fewer active parameters and the stated local-hardware path may start with Step 3.5 Flash. Neither model was installed or tested for this comparison.

VendorClue read the project pages for MiMo-V2-Flash on 7 November 2026 and Step 3.5 Flash on 7 November 2026. Terms and releases can change, so check the current pages before you choose.

Sources

MiMo-V2-Flash: mimo-v2-flash-project · Evidence: documentation · Not checked

Step 3.5 Flash: step-3-5-flash-project · Evidence: documentation · Not checked

Next step

Tools in this guide

All AI models tools
All guides