MiMo-V2.6-Flash-RL vs Qwen3.5-4B

MiMo-V2.6-Flash-RL 59

Model card 14, Intended use, the architecture, the parameters, the context and the languages (in the metadata) are on the card. Training data 5, A description of the training method with no data sources, sizes or cut-off. Last change 20, the last commits, on 28 October 2026, updated the card, 13 days before this check. Responsiveness 5, Of the 13 most recent discussions that aren't pull requests (21 September to 6 November 2026), 3 have a reply from a XiaomiMiMo member, which is a few. Version policy 5, MiMo-V2.5 stays in its own repositories and V2.6 came out in new ones, with no dated note on the older cards that we read. File health 10, Config, template and card agree on the architecture and the 1M context, and two early faults (invalid JSON in the draft model's config, preprocessor values) were fixed in commits that say so. 14 + 5 + 20 + 5 + 5 + 10 = 59.

Qwen3.5-4B 34

The card covers five of the seven items and misses limitations and a cut-off (14). Training data is a general description without a cut-off (5). The files last changed on 2 March 2026, 245 days before 2 November, which is within a year (5). No Qwen member has answered any of the 21 discussions that aren't pull requests, four of them opened since 21 October (0), and the GitHub issues couldn't be counted. Earlier 4B models stay in their own repositories without dated notes (5), and two pull requests fixing the template and adding generation_config.json wait unanswered (5). 14 + 5 + 5 + 0 + 5 + 5 = 34.

MiMo-V2.6-Flash-RL 65

Memory 5, the smallest build at 4 bits or better is ggml-org's MXFP4 GGUF, 168.93 GB with its projector, and 168.93 GB × 1.2 = 202.7 GB, within 256 GB. Xiaomi's own files are 177.74 GB. ggml-org's Q2_K is below 4 bits and isn't read. Runtimes 20, Released vLLM 0.33.0 and SGLang 0.5.23 have the mimo_v2 model and the mimo parsers, Transformers 5.22.2 has a mimo_v2_flash model, and llama.cpp b12242 has the architecture, with a GGUF from its own organisation. Builds 10, one precision from Xiaomi, the 4-bit files it releases. Terms 15, MIT in the card's metadata, an OSI licence. Download 10, the Hub API reports gated false on 10 November 2026, so the files download with no account and nothing to accept. Hosted 5, the card says it is on Xiaomi's MiMo Open Platform API and on OpenRouter, and Novita is live for it on Hugging Face. 5 + 20 + 10 + 15 + 10 + 5 = 65.

Qwen3.5-4B 85

Memory earns the full 30 because Ollama's own Q4_K_M build is 3.3 GB, about 4 GB with the fifth added, though Alibaba itself publishes no quantised 4B, so builds scores 0. Runtimes reach the 25 cap on vLLM's stable release, Transformers 5.2.0, llama.cpp, Ollama's library and LM Studio's catalogue. Apache 2.0 with nothing added gives the full 15 for terms, the ungated repository the full 10 for download, and two providers (featherless-ai, live on Hugging Face on 2 November, and Doubleword's own list per the 21 October check) the 5 for hosted too. 30 + 25 + 0 + 15 + 10 + 5 = 85.

MiMo-V2.6-Flash-RL 85

Chat template 20, chat_template.jinja in the repository (3,868 characters), with a copy in tokenizer_config.json. Tool-calling format 25, the card names the parser to use in two runtimes: the mimo tool-call parser in SGLang and in vLLM, with auto tool choice, and both ship it in a release. Trained tool calling 20, the card describes reinforcement learning on coding and general agent tasks across several harnesses, and the template renders every call in a turn. Structured output 0, Nothing on the card about JSON or schema-constrained output. Reasoning control 10, the template has an enable_thinking switch and keeps reasoning_content, and the card names the mimo reasoning parser for SGLang and vLLM, which returns the reasoning apart from the answer. Context 10, 1M tokens per the card's model summary, and 1,048,576 positions in config.json. 20 + 25 + 20 + 0 + 10 + 10 = 85.

Qwen3.5-4B 70

The template ships in the repository (20) and the model was trained for tool use with several calls per turn in the template (20). Tool-calling format gets 10, because the card names the qwen3_coder parser but leaves the declaration and call format to the template. Thinking turns off per request and comes back in think tags split by the qwen3 parser (10), and 262,144 tokens is native (10). Structured output scores 0, as nothing documents JSON output for the model.

MiMo-V2.6-Flash-RL 45

File safety 5, the weights are safetensors, but the dflash folder's mask_embedding.pt is a pickle beside them (15). Provenance 15, the files sit in the XiaomiMiMo organisation, which mimo.xiaomi.com links to per the Distill listing (10). Safety evaluation 0, no safety results on the card. Agent guidance 5, General cautions only. Licence risk 15, MIT has no use policy, threshold or change clause. Security programme 5, Xiaomi runs a company security centre for reporting flaws, as read for the Distill listing (5). 5 + 15 + 0 + 5 + 15 + 5 = 45.

Qwen3.5-4B 55

Safetensors only with no remote code gives file safety its 25, and files in the Qwen organisation the publisher's GitHub links to, each with a hash on the Hub, give provenance 15. The Apache 2.0 licence carries none of the deductions, so licence risk keeps 15. Safety evaluation, agent guidance and a security programme score 0, since the card has no safety results or cautions and no reporting route was found. The Qwen3.5 blog and Alibaba's security response centre couldn't be read, so those two are scored as absent and left open.