MiMo-V2.6-Flash-MOPD Mixed-Quant GGUF

One mixed-precision variant of XiaomiMiMo/MiMo-V2.6-Flash-MOPD, pinned to 2479e2d0029eca9a34cc7e7f55a121925f81908e.

The tensor recipe matches the RL release: routed expert gate/up IQ2_XXS, expert down IQ2_XS, shared attention/dense/embedding/output and three embedded MTP blocks Q8_0, routers/norms/sinks F32. Input vision/audio components remain BF16/F32. The separate five-layer DFlash draft is Q8_0/F32, with its original mask embedding and configuration.

Artifact inventory

Main backbone: 88.778 GB across four shards. Six GGUF weight files including media and DFlash total 93.092 GB (86.699 GiB). The main tensor payload is exactly 88,771,782,144 bytes, matching the RL recipe. Calibration and reproduction files are additional.

Exact SHA256 values are in SHA256SUMS and artifact-manifest.json.

Download and text inference

Download the complete variant and verify its six weight files:

hf download Baekpica/MiMo-V2.6-Flash-MOPD-Mixed-Quant-GGUF \
  --include 'MQ-IQ2-XXS-XS-Q8-MM-BF16/*' \
  --local-dir ./MiMo-MOPD
cd MiMo-MOPD/MQ-IQ2-XXS-XS-Q8-MM-BF16
sha256sum -c SHA256SUMS

Keep all four main shards in the same directory and pass the first shard to the loader. The calibration build uses llama.cpp commit 58367713a6935c0810103378144008df32e3d5db; its build and conversion details are in reproduction/README.md. The following text invocation uses the embedded upstream chat template and disables thinking:

/path/to/llama.cpp/build/bin/llama-cli \
  -m MiMo-V2.6-Flash-MOPD-MQ-IQ2-XXS-XS-Q8-00001-of-00004.gguf \
  -ngl 999 -c 8192 -b 512 -ub 512 -n 128 -st --temp 0 \
  --chat-template-kwargs '{"enable_thinking":false}' \
  -p 'What is 17 plus 25? Answer briefly.'

The local checks use three 96-GB RTX PRO 6000 Blackwell GPUs and a container RAM limit of 272,000,000,000 bytes. File size alone does not specify device-memory requirements: the runtime also allocates KV cache and work buffers. Embedded MTP and the separate DFlash package are retained, but their speculative decoding quality and speed are not qualified by the plain-decoding checks.

Image/video/audio use the native input contract described below. The heldout backbone comparison inserts native embeddings; it does not validate generic --image, --audio, or --video preprocessing for every modality. The BF16/F32 media file contains input-modality components, not a qualified audio-output serving path.

DGX Spark performance

Measured on 2026-09-30 with this MOPD artifact (MQ-IQ2-XXS-XS-Q8-MM-BF16, revision 16965fb66e6ab3290e6cd20ddbfa30adda692764) on one NVIDIA GB10 / DGX Spark. The ds4-dfm-rs comparison uses baseline 7a78fcd8 and retained 2db2fc38 from Prefill PR #68. The retained build contains direct token-compact Q8 input and pipelined Down worklists; their scoped 8K adoption evidence is in the runtime report.

Prefill here measures the newly added 2,048-token suffix at each context frontier. It is not a cold prefill of the complete 64K prefix. Decode generates 128 greedy tokens, excluding EOS, at every frontier. K means 1,024 tokens. Lines are per-frontier medians of three fresh processes per build; bands show the observed minimum–maximum.

At 64K, the retained build measures 663.48 tok/s incremental prefill (+0.68% versus baseline) and 17.80 tok/s greedy decode (-0.34%). Its 64K rates retain 59.73% of the 2K prefill rate and 70.33% of the 2K decode rate. These are this fixture's measurements.

Conditions: CUDA 13.3.73, driver 615.71.09; configured GPU clock range 300–2200 MHz, observed busy clocks 2184–2190 MHz. One shared VMM weight owner and one worker at a time; prefill capacity 4,096, allocated context 65,665; embedded MTP and DFlash speculation disabled. Baseline/final process order rotates B/F, F/B, B/F. Each sweep has a separate complete 8K/128 warmup process, then one fresh session progressing through 32 frontiers. The benchmark restores the pre-generation session snapshot before computing the next suffix. Loading, tokenization, snapshot I/O and proof-file I/O are outside phase timers.

The raw-text fixture is speed-bench/promessi_sposi.txt, SHA256 f53e0d80cb2d4492d24ebd63c7000c397b16ae70f9bf09b3763e5d8323ec209f. All 192 measured frontiers have 152,576 finite logits and 128 generated token IDs; both logits and token streams match byte-for-byte across all six processes at every frontier. This checks the retained changes on the measured text workload; it does not establish general model quality, request TTFT, or concurrent/multimodal 64K serving qualification.

Evidence: protocol and build hashes, curve values and correctness hashes, raw CSVs baseline 1, 2, 3, and retained 1, 2, 3.

Calibration

MOPD backbone activations are collected from an original-representation reference: source MXFP4 expert blocks are repacked without further quantization; source FP8 dense weights are expanded to BF16. This is not an original full-BF16 checkpoint. RL activation statistics and RL media embeddings are not reused.

The compact text corpus contains 122 records and 234,844 MiMo tokens. It combines 100,291 tokens of prior reasoning, code, general and retrieval fragments, 68,827 tokens of interactive agent prefixes, and 65,726 tokens from resolved code trajectories. Structured messages and tool arguments use the MOPD chat template and tokenizer; Unicode normalization follows its HF tokenizer. Trajectory selection removes duplicate calls within one turn and consecutive identical tool-call batches. This filter is a sampling heuristic, not a complete semantic repetition detector.

Media calibration selects 48 chart/document images, 32 speech records, 24 real video clips and 24 joint audiovisual inputs. Prepared media records contain 63,567 decoder positions, including 56,889 inserted media embedding positions. Together with 234,844 text positions, the compact corpus totals 298,411 positions: approximately 78.7% text-only records and 21.3% media-containing records. Counts of records are not modality token percentages; exact category totals are in prepared-media-summary.json. The pool is chart/document and speech focused; it does not provide broad UI-agent or acoustic-event coverage.

Media follow this actual computation path:

source image/video/audio -> MOPD native encoder/projector -> projected embeddings + MiMo tokens -> LLM backbone -> activation statistics

The collector requests outputs for every calibration position so the runtime cannot prune final-layer FFN rows. The per-layer observation audit requires all 47 routed layers to account for every submitted position. The collector submits float embeddings to the LLM graph. Media filenames and placeholder tokens alone are not calibration inputs. Production processing retains ImageNet normalization, two-frame temporal patches, explicit video boundaries and timestamps, the original audio tokenizer and local encoder, and per-frame-pair audiovisual interleaving. Every prepared audio token in an AV input must be covered exactly once.

The technical report, §5.6 and tool repetition analysis motivated agentic history prefixes and actual multimodal activations. Their private training mixture is not public. This is a practical calibration proxy, not a reconstruction of the exact MOPD teacher distribution or proof that quantization preserves its improvements.

Collected importance matrix

All 122 text records (234,844 positions) and all 128 native media records (63,567 positions) completed with finite final logits. For each of the 47 routed layers, every gate/up/down projection accounts for exactly eight expert selections per submitted position, including the final layer. The two raw matrices were merged by adding their sums and counts; all 574 resulting tensors were checked against that operation.

The final matrix observes 12,029 of 12,032 layer/expert pairs. Layer 7 experts 13, 89 and 184 remain unobserved. Their genuine zero counts are retained; the pinned quantizer assigns uniform column weights to those experts. No observations are fabricated and no layers are promoted. Final imatrix SHA256: f37aa247d5f0fa6b3bc021e6b0fb87a39f4fd6d6354a0f6742df4047ede600cd.

Text and media routing frequencies differ in the collected data; the largest layer-level Jensen–Shannon divergence is 0.2285 nats at layer 7. The median relative L2 change in mean-square inputs from the 60-text-record snapshot to the final text matrix is 0.0517 (90th percentile 0.3455). These are descriptive diagnostics, not a convergence guarantee or task-quality score. Full observations are in the variant's imatrix-diagnostics.json.

Scope of validation

HF CLI verified all 154 pinned source files (177,765,031,074 bytes). Reference conversion checks covered all 36,096 routed expert matrices using deterministic sampled rows, plus TP=4 QKV ordering/scales. The final main GGUF audit passed all 508 tensor names, shapes and quantization types, exact preservation of 248 F32 tensors, tokenizer/architecture metadata, shard boundaries, and sampled finite expert values. Q8 matrix samples had relative L2 error below 0.03. The BF16/F32 media file passed finite-value checks; one-image native/GGUF projector parity measured cosine similarity 0.999682 and relative L2 error 0.025522. These are conversion checks, not general accuracy guarantees.

Both reference and mixed models returned 42 through the actual pinned llama-cli for the arithmetic smoke test. The separate 16-record comparison used greedy decoding with a 64-token output limit and full-vocabulary first-token logits:

The three exact-answer-scored text prompts (arithmetic, Korean counting, code) matched their expected answers in both models. Manual inspection also found the expected tool-stop result, translation, and JSON values in both; JSON remained fenced rather than a bare object. Mean word error rate on the two speech samples was 0.04167 for both models. Image whole-response exact match was 0/4 for both because responses were verbose, with three responses per model reaching the token limit. Four video/AV responses per model also reached that limit, and these samples have no task-accuracy labels.

The mixed model showed an isolated stray Chinese character in one English chart response and an analysis-style opening in one AV response. Video/AV first-token divergence was larger than text/audio divergence. These observations are disclosed rather than treated as evidence of parity; this small comparison does not establish preservation of MOPD's post-training gains. The complete generated responses and metric limitations are in holdout-comparison.json, with the release decision in release-review.json. No RL DGX Spark performance measurements are claimed for these weights.

A small heldout comparison uses six text prompts and ten media inputs, including two correlated video/AV pairs. It is separate from all importance-matrix collection. Native source media embeddings are shared by the reference and mixed language-model checks; this does not qualify every deployment runtime or establish general task accuracy.

Reproducibility and licenses

The release includes the exact per-tensor recipe, final imatrix, source revisions, calibration selection scripts and provenance, conversion/collection code, checksum manifest and actual validation reports. Original calibration media and full source text are not redistributed. Dataset licenses remain with their respective sources. Model weights inherit the upstream MIT license.

Sources: Inkling calibration collection, Nemotron-Agentic-v1, OpenHands sampled trajectories, ChartQA, DocVQA, LibriSpeech, and FineVideo.