Known-issues sheets for local language models Rev 0.1.0 · reviewed every 7 days
llm-picker

Model datasheet · MiMo-V2.6

MiMo-V2.6-Flash (MOPD)

Xiaomi's 309B MoE with 15B active, as the MOPD checkpoint: the 27 September follow-up to the RL release, retrained to stop the repeated tool calls the RL files were reported for. Xiaomi ships the experts in MXFP4, so the MXFP4 GGUF (167-175GB, depending on how the other layers are stored) is the model as released and there is nothing bigger worth downloading; the only smaller stock file is a 126GB Q2_K. 39 of 48 layers use a 128-token sliding window, so the cache stays small: the 9 global layers hold 23KB per token, about 22GB at a million tokens. That figure is computed from config.json, with keys 192 wide and values 128 wide, and is not a measured llama.cpp log. Every benchmark on this sheet was published for the RL checkpoint, not this one. Mainline mlx-lm cannot load this architecture yet, so no MLX link is given.

Vendor
Xiaomi
Architecture
MoE · 309B total / ~15B active
Context
1,048,576 tokens
License
MIT
Released
2026-09-27
Vision
Yes

→ GGUF (llama.cpp / LM Studio / Ollama)

Architecture, from its own config.json

48 layers · 9 full attention · 39 sliding window

64 attention heads · mixture of experts, 15B of 309B active per token across 8 experts · vocabulary 152,576

Where the memory goes: weights by band above, plus a KV cache of 326MB at 8K · 866MB at 32K · 3GB at 128K, plus the runtime's own working memory — dominated by the output projection over that vocabulary.

Inside one layer, from its own tensors

9 × full-attention layer

input_layernormattentionqkv_proj — q, k and v in one tensorsoftmax(q·kᵀ/√d) · v64 query heads over undefinedkey/value headso_projpost_attention_layernormmixture of expertsrouter picks 8 of 256 expertseach 2,048 wide: gate·up → down

39 × sliding-window layer

input_layernormattentionqkv_proj — q, k and v in one tensorsoftmax(q·kᵀ/√d) · v64 query heads over undefinedkey/value headso_projpost_attention_layernormmixture of expertsrouter picks 8 of 256 expertseach 2,048 wide: gate·up → down

The green line down the right of each half is the residual: it leaves the spine above the norm, skips the operation entirely and rejoins at the +. Where attention is drawn open, q, k and v are scaled to their real output widths — a model that shares key and value heads across many query heads shows one wide box beside two narrow ones.

Around the stack: an embedding of 152,576 × 4,096, a final norm, and a separate output projection. The first 1 layer use a dense feed-forward instead of the experts.

On the norm names: despite what it is called, post_attention_layernorm is drawn after the residual add here, which is where this family applies it — it normalises the input to the feed-forward, not the output of attention. Models that carry a separate pre_feedforward_layernorm use it the other way, and are drawn that way.

Also in every block and not drawn, because their tensor names do not fix where they belong: e_score_correction_bias, attention_sink_bias. Guessing an order is the one thing this diagram will not do.

§1 Characteristics by quantization band

Q2–Q3 — 2–3 bit

126GB weights · e.g. Q2_K + KV: 326MB at 8K · 866MB at 32K · 3GB at 128K
Coding 6/10
data checked Oct 2026
  • quant-degradation-community — editorial band derivation @ Q2_K · aggregated · 2026-10-04

    Editorial derivation: two points down from the MXFP4 release. ggml-org's Q2_K keeps the expert down projections at MXFP4 and takes gate/up to Q2_K. No measurement of that file was found; the nearest is AesSedai's table for their own mixes of this checkpoint, where 3.0 bits per weight costs 5.5% perplexity and mean KLD 0.17, and 3.5 bits costs 2.0% and 0.13.

Tool calling / agentic 6/10
data checked Oct 2026
  • quant-degradation-community — editorial band derivation @ Q2_K · aggregated · 2026-10-04

    Editorial derivation: two points down, on the same perplexity and KLD evidence as coding. Agent loops are where low-bit damage usually shows first, and no one has published a tool-calling run at this size.

Q4–Q5 — 4–5 bit

175GB weights · e.g. mxfp4 + KV: 326MB at 8K · 866MB at 32K · 3GB at 128K
Coding 8/10
data checked Oct 2026
  • vendor-model-card — Terminal-Bench 2.1 = 87.6 @ mxfp4 · vendor-reported · 2026-10-04

    RL CHECKPOINT, not MOPD. Published on the MiMo-V2.6-Flash-RL card; the MOPD card has no benchmark table. The only vendor statement linking the two is the tool-call blog post (https://mimo.xiaomi.com/blog/mimo-v2-6-tool-call-repetition): after MOPD, 'performance on the broader benchmark suite held steady'. Xiaomi does not say which reasoning mode its own model ran in; the technical report says only that baselines with configurable effort ran at max. Xiaomi ships the experts in MXFP4 (store_dtype mxfp4 in config.json, MXFP4 QAT per the technical report), so the release precision is recorded here.

  • vendor-model-card — DeepSWE v1.1 = 67.9 @ mxfp4 · vendor-reported · 2026-10-04

    RL CHECKPOINT, not MOPD; same caveat as above. Above GLM-5.3-Flash's 63.4 on Z.ai's run of the same benchmark. Scored one point below that model's 9 anyway, because these figures belong to a sibling checkpoint and the only bridge to this one is a sentence, not a table.

Tool calling / agentic 8/10
data checked Oct 2026
  • vendor-model-card — AutomationBench v1.0.6 = 52.3 @ mxfp4 · vendor-reported · 2026-10-04

    RL CHECKPOINT, not MOPD. The RL checkpoint is the one with the reported tool-call repetition loop, which this benchmark did not catch; MOPD was trained to fix that, and Xiaomi says the wider suite 'held steady'. Reasoning mode not stated.

  • vendor-model-card — Toolathlon-Verified = 73.6 @ mxfp4 · vendor-reported · 2026-10-04

    RL CHECKPOINT, not MOPD. Same table: Agents' Last Exam 27.6, OSWorld-Verified 80.8. The GDPval-AA row on that card is an Artificial Analysis metric and is not recorded here.

§2 Known issues & what fixes them

ISSUE-01 moderate Q2–Q3 / Q4–Q5 · tool-calling, coding

Check which files you downloaded. The earlier RL checkpoint, which ggml-org publishes under an almost identical name ending in -RL, was widely reported for issuing the same tool call over and over: one OpenCode user said it searched 'thousand times', another described a cascade of about 35 video calls, sixteen of which downloaded the same file. Xiaomi acknowledged it, measured it at 1.02% of responses in OpenCode for Flash-RL, and released this MOPD checkpoint on 27 September to fix it. Their own chart shows the rate falling substantially, not to zero, so some repetition may remain here too.

No fix
No known fix — choose a different model [strong]

If your files say -RL, replace them with the -MOPD ones (ggml-org/MiMo-V2.6-Flash-MOPD-GGUF). One of the users who reported the loop says the MOPD release fixes it. Nothing on the prompt side fixed it on the RL checkpoint.

Evidence · 4 sources · community-consensus
ISSUE-02 moderate Q2–Q3 / Q4–Q5 · coding, tool-calling

Conversion support reached mainline llama.cpp on 22 September and DFlash drafting on 30 September, so the support is under two weeks old. In the first week the MTP drafter files failed to load until they were re-uploaded and master was fixed (25-27 September), and on 28 September ggerganov reported a DFlash crash of his own. An open report describes tool-call arguments turning into malformed JSON late in long agent sessions. It was filed against the RL checkpoint over RPC at 262K context, so whether MOPD shows it is not known.

Fix
Update your runtime [partial]

Use a current llama.cpp build and re-download the drafter files if they fail to load. The people who hit the loading failure said the model ran fine without the drafter.

No fix
No known fix — choose a different model

Nothing on the prompt side fixes a backend bug.

Evidence · 4 sources · community-consensus
ISSUE-03 mild Q2–Q3 · coding, tool-calling

Xiaomi trained the experts in MXFP4 and ships them that way, so a Q2_K file is not trimming spare precision. It is cutting into the model as released. AesSedai's measurements of their own mixes of this checkpoint put 3.0 bits per weight at 5.5% worse perplexity than the MXFP4 file and 2.0 bits at 27%. No one has measured ggml-org's Q2_K directly.

Fix
Move up a quant band [strong]

The MXFP4 file is 41-49GB larger than the 126GB Q2_K and is the full model. There is no step in between from ggml-org.

Evidence · 2 sources · anecdotal
ISSUE-04 mild Q2–Q3 / Q4–Q5 · coding, tool-calling, math, factuality, instruction-following

Every score on this sheet comes from the RL checkpoint's card, because the MOPD card publishes no table. Xiaomi's only statement linking the two is that performance on the broader suite 'held steady' after MOPD. That table covers agentic coding, tool use and security only, so math, factuality, instruction following, long context and writing are unrated because nothing was published, not because the model is known to be weak at them.

No fix
No known fix — choose a different model

Not a defect. It is a caveat, and the blank cells are honest.

Evidence · 2 sources · anecdotal