[!WARNING] Experimental low-bit release. The routed experts, which hold most of the weights, are stored at about 1.6 bits per weight (2.6 on the 32 layers with an 8-bit residual stage), far below common 4-bit GGUF quants. You may see failures the source model rarely shows: reasoning that loops or never reaches an answer, text that breaks mid-character (shown as
�) or switches language mid-word, and weaker answers in general. This is exploratory quantization research, not a drop-in replacement for the source model.The files can change. This line of conversions is still being researched, so any file here may be replaced by a new revision. Replacements so far: none. To keep a fixed build, download a pinned commit:
hf download Baekpica/MiMo-V2.6-Flash-MOPD-VQ-GGUF --revision <commit>.Problems or odd output? Feel free to post them in the Community tab, even as rough notes. The prompt, sampling settings and runtime commit help most.
Xiaomi's MOPD upgrade of MiMo-V2.6-Flash-RL — the same 309B native omnimodal MoE (15B active), post-trained by multi-teacher on-policy distillation (MOPD2) and a short specialized-teacher run against tool-call repetition — with its full 1,048,576-token context, its vision/video tower, its audio tokenizer and audio encoder, its DFlash draft model and embedded MTP heads, sized for one 128 GB DGX Spark. The 166 GB checkpoint (MXFP4 experts, FP8 attention) becomes 98.5 GB of GGUF parts; with the 1M-token KV cache, both media sidecars and the drafter loaded at once, 111.1 of the 117.2 GiB a GB10 offers stay in use. Routed experts are 8-dimensional vector-quantized codes — 13-bit everywhere, plus an 8-bit residual stage on 32 layers — fitted with an importance matrix from 4.13M calibration tokens (text, chat, tools, images, video, audio) run through the MOPD checkpoint itself. Measured against the MOPD original, teacher-forced on held-out chats: KL 0.083 and the same next token 92.2% of the time on assistant turns (the original against itself, batched differently: KL 0.017, 98.1%).
MOPD changes weights only: the architecture, configuration, tokenizer and chat-template output are the RL checkpoint's. This release therefore applies the recipe of MiMo-V2.6-Flash-RL VQ GGUF unchanged — the same formats, the same 32 residual layers, the same part layout and byte sizes, the same calibration set — with statistics and codebooks computed from the MOPD weights. The predecessor, MiMo-V2.6-Flash-MOPD Mixed-Quant GGUF, used IQ2_XXS/IQ2_XS scalar experts; this release takes the VQ route ds4-dfm-rs runs for Beam-1.0, DeepSeek-V4.1, GLM-5.3, MiniMax-M3.1 and the MiMo-V2.6-Flash-RL releases.
| Source | XiaomiMiMo/MiMo-V2.6-Flash-MOPD at 5220c3f5 (routed experts MXFP4, attention/dense FP8 128×128 blocks, o_proj/embeddings/head BF16), the MOPD post-training of MiMo-V2.6-Flash-RL f657ba2f |
| What MOPD changed | the language model's weights: routed experts, attention, dense FFN, norms, token embedding and head. Unchanged from RL (identical source bytes): the vision tower, the audio tokenizer and audio encoder, the DFlash draft model and the embedded MTP blocks |
| Model | 48 layers: 9 global-attention (GA: 0, 5, 11, 17, 23, 29, 35, 41, 47; 64 query / 4 KV heads) and 39 sliding-window layers (window 128, 64 / 8 KV heads, learned attention sink); head dims QK 192 / V 128, partial RoPE (64 of 192 dims, θ 1e7 GA / 1e4 SWA); layer 0 dense FFN 16,384; layers 1–47: 256 routed experts, top-8, sigmoid + correction bias, no shared expert, expert FFN 2,048; vocabulary 152,576; 302.8B routed-expert weights |
| Recipe | mimo26f-vq8-r13x47-r8x32-q8-fp8kv-embssd-v1 (recipe.json), the RL releases' recipe |
| Routed experts | 8-dim VQ, one E4M3 codebook per layer shared by all 256 experts and gate/up/down, one F16 gain per row. 13-bit (K = 8,192) on all 47 MoE layers; an 8-bit residual stage (K = 256) on 32 layers (R13+8): 5, 8, 11, 13, 17, 19–25, 27–30, 32–47 — the 8 GA MoE layers first, the other 24 from the end-to-end placement measurement of the RL release. 1.63 bpw on 15 layers, 2.63 on 32, 2.31 on average |
| Attention qkv / o, dense FFN (layer 0), output head | Q8_0 (FP8 sources dequantized exactly first; the TP=4-interleaved fused qkv is stored as [q | k | v]) |
| Kept at source precision | router (BF16) and correction bias (F32), every RMSNorm, attention sink biases |
| Token embedding | BF16 source bytes in its own part, read in rows from SSD (one 8,192-byte row per token) |
| Vision / video tower | MiMo ViT (28 layers, 24 window + 4 global) + 2×2 merger, Q8_0 sidecar, resident (0.72 GiB); byte-identical to the RL release's part |
| Audio | AudioTokenizer encoder (24 layers, 20 RVQ codebooks) + audio patch encoder (6 layers) + projector, Q8_0 sidecar, resident (0.61 GiB); RVQ codebooks F32; byte-identical to the RL release's part |
| Speculative decoding | DFlash draft model (5 SWA-1024 layers, block 8, mask embedding) Q8_0 1.45 GiB; the three embedded MTP blocks Q8_0 0.98 GiB as the alternative drafter; both byte-identical to the RL release's parts |
| Context | 1,048,576 tokens, the model's own; one sequence on one DGX Spark (serial lane) |
| KV contract | GA layers: FP8 E4M3 + F16 scale per KV-head row, resident, 1,296 B per token and GA layer, 11.39 GiB at 1M; SWA layers: 128 + chunk − 1 rows (BF16) |
| Chat template | Original MOPD chat_template.jinja, unchanged (<think> reasoning, enable_thinking, XML tool calls, image/video/audio placeholders) |
| Fidelity (release bytes vs the MOPD original, assistant turns) | KL 0.083, same top-1 92.2% (the RL release, same recipe, vs its original: 0.088 / 92.2%; details) |
| Runtime | ds4-dfm-rs, family mimo2_vq, the RL releases' runtime |
MOPD changes weights only, and this release keeps the RL releases' formats, shapes, part layout and byte sizes, so it
runs on the same ds4-dfm-rs mimo2_vq kernels at the same speed. The numbers below are those recorded for the
MiMo-V2.6-Flash VQ releases: measured 2026-10-31 with ds4-dfm-rs 2fa8c08e on the UNCENSORED release
(RL card).
One DGX Spark / GB10, CUDA 13.4.92, driver 615.71.10; operator-managed 300–2,200 MHz range, observed busy SM
2,177–2,190 MHz. Three fresh ds4-bench processes per variant; each of 32 frontiers from 2K to 64K
adds 2,048 prompt tokens of speed-bench/promessi_sposi.txt and generates 128 forced greedy tokens, without
drafting and with two embedded MTP drafts; 4,096-row prefill cap; FP8 full-layer K/V. Receipts and raw CSVs.
| Median of three run means | Prefill tok/s | Decode tok/s |
|---|---|---|
| No drafting, 2K–64K | 1,294.9 | 23.96 |
| No drafting, 2K–32K | 1,400.4 | 24.46 |
| Two MTP drafts (server default), 2K–64K | 1,292.2 | 26.35 |
| Two MTP drafts, 2K–32K | 1,398.1 | 25.48 |
| Frontier | 2K | 8K | 16K | 32K | 48K | 64K |
|---|---|---|---|---|---|---|
| Prefill tok/s | 1,492.2 | 1,472.3 | 1,402.8 | 1,293.4 | 1,185.8 | 1,106.1 |
| Decode tok/s, no drafting | 24.86 | 24.78 | 24.58 | 24.04 | 23.43 | 23.04 |
| Decode tok/s, two MTP drafts | 23.18 | 25.04 | 27.91 | 24.38 | 26.23 | 28.40 |
Drafting on chat prompts, one process each, 512 greedy tokens, two MTP drafts (RL-family measurement):
| Prompt | No drafting | Two MTP drafts | Tokens per cycle |
|---|---|---|---|
| Write a Rust LRU cache (105 tokens) | 24.96 | 35.96 | 2.81 |
| Comment a 156-line C file (2,365 tokens) | 25.05 | 36.29 | 2.93 |
| Explain speculative decoding, Korean (152 tokens) | 24.92 | 25.40 | 2.01 |
MTP decode follows draft acceptance, which depends on the weights: MOPD keeps RL's MTP blocks unchanged while its trunk moved, so its acceptance (and the MTP rows above) can differ from the RL family's; the no-drafting rows depend only on shapes.
# native 1,048,576 context, vision + audio, two embedded MTP drafts (serial lane)
ds4-server --cuda -m MiMo-VQ-DIR
# DFlash drafter instead of MTP
ds4-server --cuda -m MiMo-VQ-DIR --mtp MiMo-VQ-DIR/dflash-q8_0.gguf --mtp-mode on --mtp-draft 7
# two text banks with partial prefix reuse and disk KV
ds4-server --cuda -m MiMo-VQ-DIR -c 262144 --max-seqs 2 --mtp-mode off \
--prefix-reuse partial --kv-disk-dir KV-DIR
Plan (ds4-server --check-config, RL-family releases) |
Lane | GiB of 117.6 |
|---|---|---|
| default: 1M, vision + audio, MTP 2 | serial | 109.79 |
| 1M with the DFlash drafter | serial | 110.27 |
| 512K, one text bank + serial media, no drafting | banks | 111.77 |
| 256K, two text banks + serial media, partial reuse | banks | 109.89 |
The vision and audio parts load with the model (no --vision). Images, video and audio run on the serial lane;
text banks keep their own sequences, so a bank plus the media session at 1M (123 GiB) is refused.
MiMo keeps the full history only in its 9 global-attention layers; the 39 sliding-window layers attend to the last 128 tokens. Its KV is therefore small enough to stay resident at 1M — no SSD tier is needed — once the GA cache is stored in FP8 instead of BF16:
| GA cache format | Bytes per token, 9 GA layers | At 1,048,576 tokens |
|---|---|---|
BF16 (existing mimo2 runtime) |
23,040 | 22.50 GiB |
| FP8 E4M3 + F16 row scale (this release's contract) | 11,664 | 11.39 GiB |
On the RL weights the FP8 cache gives assistant-turn KL 0.0185, inside the original's own batching noise (0.020). The memory saved goes into expert precision.
A GB10 CUDA block can use at most 101,376 bytes (99 KiB) of shared memory, and the routed-expert kernels keep the layer's whole codebook there while they decode every selected expert:
| Recipe | Base book | Residual book | Total |
|---|---|---|---|
| R12 | 32 KiB | — | 32 KiB |
| R13 | 64 KiB | — | 64 KiB |
| R13+8 | 64 KiB | 2 KiB | 66 KiB |
| R13+10 | 64 KiB | 8 KiB | 72 KiB |
| R14 | 128 KiB | — | 128 KiB (does not fit) |
R13 is the widest base book that fits; the 8-bit residual stage adds only a 2 KiB book and cuts a layer's weight error about threefold (MOPD experts, relative MSE on layers 12–47: 13-bit 0.139–0.161, R13+8 0.048–0.060). The remaining ~33 KiB holds activations, MMA tiles and staging. MiMo's expert geometry (hidden 4,096, expert FFN 2,048, top-8) is exactly GLM-5.3 VQ's:
| Expert kernel stage (shared memory per block) | R13 | R13+8 |
|---|---|---|
| Decode gate/up (book + 4,096-wide BF16 input row) | 73,728 B | 75,776 B |
| Decode down (book + all eight experts' 2,048-wide BF16 rows) | 98,304 B | 100,352 B |
| Prefill MMA (MiniMax/GLM tiles) | 96,768 B | 98,816 B |
Against the 117.2 GiB a GB10 reported available to ds4-server --check-config (Beam-1.0 VQ), with every optional
component loaded at once (part bytes, identical to the RL release's layout):
| GiB | |
|---|---|
| Routed experts, 47 layers × 13-bit (codes, gains, codebooks) | 57.47 |
| 8-bit residual stage on 32 layers | 24.01 |
| Attention qkv/o, dense FFN, head (Q8_0); router, norms, sinks | 5.35 |
| DFlash draft model (Q8_0) | 1.45 |
| Vision sidecar (Q8_0) | 0.72 |
| Audio sidecar (Q8_0) | 0.61 |
| GA KV, 9 layers × 1M tokens, FP8 | 11.39 |
| SWA rings (BF16, 4,096-row chunks) and drafter rings | 0.85 |
Graph scratch, 4,096-row prefill chunks (mimo2 planner) |
2.83 |
| Media workspace for the largest image (32,768 patches) | 2.38 |
| Floor kept free by the server | 4.00 |
| Resident at 1M, everything loaded | 111.06 of 117.2 |
| On SSD, read on demand: token embedding (BF16) | 1.16 |
Measured on GB10 for the RL-family releases (2026-10-30): ds4-server --check-config plans 110.27 GiB for this
case (1M, the DFlash drafter, vision, audio and the media workspace) and 109.79 GiB for the default with two MTP
drafts, of 117.6 GiB available.
The RL release's calibration set, unchanged: 4.13M tokens; held-out sets (27 text cases, 63,337 tokens; 50 native text/image/video/audio inputs, 21,288 tokens) reserved first. The MOPD tokenizer is the RL one, and its chat template differs from RL's by one blank line between template blocks, which whitespace control removes from the output, so the token rows are the ones MOPD's own template renders. The media embeddings come from the vision tower and audio encoders, which MOPD left unchanged.
| Part | Rows | Tokens | Content |
|---|---|---|---|
| Text and chat | 1,579 | 3,661,171 | English 39.0%, code 14.0%, Korean 13.9%, Chinese 13.2%, Japanese 7.4%, Spanish 3.9%, ds4 agent prompts 2.9%, French/German/Indonesian/Russian/Vietnamese; raw text 38.0%, thinking chats 33.9%, non-thinking chats 18.6%, tool/agentic 9.5%; no domain above 8.5% |
| Native media | 953 | 464,968 (418,203 media) | 486 chart/document images, 79 videos (two-frame groups, MM:SS timestamps), 309 audio clips, 79 joint video+audio clips; embeddings from the original BF16 MiMo ViT and audio tokenizer/encoder with production preprocessing |
66% of the text tokens sit in rows of 4,096+ tokens, 20.7% in 16K–32K rows and 6.5% in 32K–64K rows (longest 48,950): the sliding-window layers see 128 tokens, so long rows are what give the global-attention layers (and the experts behind them) representative statistics. Through the MOPD weights every MoE layer routes exactly 8.0 experts per token, and every layer's least-routed live expert received thousands of tokens (median over layers 21,453). As in the RL checkpoint, whose router MOPD kept, layer 7's experts 184 and 13 are never (0 of 33M routes) or almost never (10) chosen by the router; the weighting shrinks them to the layer-mean profile, so they cost the shared codebook nothing.
Teacher-forced comparison with the MOPD original on the RL release's held-out set (27 text cases and 50 native inputs; 65,223 scored positions, 24,885 of them in assistant turns), on one H200. The teacher is the MOPD checkpoint's own reference forward (BF16 math on the source weights); the student reads every routed expert from this repository's parts, uses the Q8_0 body and simulates the FP8 GA cache.
| Configuration | KL, assistant turns | Same top-1, assistant turns | KL, all positions | Same top-1, all |
|---|---|---|---|---|
| This release (MOPD): R13 ×47 + R13+8 ×32, Q8_0 body, FP8 GA K/V | 0.083 | 92.2% | 0.517 | 82.8% |
| Floor: the MOPD original vs itself, batched differently | 0.017 | 98.1% | 0.194 | 93.3% |
| RL release (same recipe) vs the RL original | 0.088 | 92.2% | 0.532 | 82.7% |
| RL floor | 0.020 | 97.8% | 0.208 | 92.6% |
On assistant turns the MOPD original's NLL is 0.774 nat/token, this release's 0.798. All positions include untrained ones (user turns, tool schemas, the text after an image or audio clip), where even the floor disagrees.
| Path | Contents | Size |
|---|---|---|
layer-LL-experts.gguf for the 15 13-bit layers (1–4, 6, 7, 9, 10, 12, 14–16, 18, 26, 31) |
routed experts of one MoE layer, GVQ1 (format version 1) | 1,313,048,192 B (layers 1–9) / 1,313,048,960 B |
layer-LL-experts.gguf for the 32 residual layers |
the same 13-bit codes plus the 8-bit residual stage (format version 2) | 2,118,611,968 B (5, 8) / 2,118,615,040 B |
body-nonexpert.gguf |
Q8_0 attention, dense FFN, head; router, bias, norms and sinks at source precision | 5,740,191,360 B |
embed-tokens-bf16.gguf |
BF16 token embedding, read in rows from SSD | 1,249,903,616 B |
vision-q8_0.gguf |
MiMo ViT + merger (the RL release's bytes) | 776,464,384 B |
audio-q8_0.gguf |
audio tokenizer encoder + RVQ codebooks, audio patch encoder, projector, speech embeddings (the RL release's bytes) | 653,195,008 B |
dflash-q8_0.gguf |
DFlash draft model with its mask embedding (the RL release's bytes) | 1,559,875,456 B |
mtp-q8_0.gguf |
the three embedded MTP blocks (the RL release's bytes) | 1,049,622,912 B |
| All 53 parts | 98,520,656,896 B | |
release.json, tensor-manifest.json, recipe.json, memory-budget.json, files.json, SHA256SUMS |
entry point, inventory with exactly-once source mapping, recipe, budget, hashes | |
config.json, generation_config.json, tokenizer*.json, vocab.json, merges.txt, chat_template.jinja, preprocessor_config.json, audio_tokenizer/config.json, dflash/config.json |
original MOPD files, unchanged | |
FORMAT.md, tools, calibration manifests, measurements, receipts |
byte format, the tools as run, calibration manifests, the measurements above, per-part receipts (incl. the MOPD byte-identity proof of the four shared parts) |
These are independent GGUF parts, not numbered GGUF splits. Standard llama.cpp, Ollama and Transformers loaders do not read this format.
LICENSE-YoungAi.mimo2_vq runtime.MIT, inherited from the pinned upstream model. Calibration datasets keep their own licenses.
The upstream model card's Use Restrictions apply to derivatives, including this quantization: no unlawful, fraudulent or malicious use, no infringement of third-party rights, no harm to individuals or groups, no military use, no malicious cyber activity, no unauthorized processing of personal data, and no autonomous execution of high-risk actions without human oversight (see the upstream README).
@misc{mimo2026v26flashmopd,
title={MiMo-V2.6-Flash-MOPD},
author={{Xiaomi MiMo Team}},
year={2026},
howpublished={\url{https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-MOPD}},
}
Have the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo Baekpica/MiMo-V2.6-Flash-MOPD-VQ-GGUF ./model-folderXiaomiMiMo/MiMo-V2.6-Flash-MOPD