j2ipper/MiMo-V2.6-Flash-MOPD-a4-x

MiMo-V2.6-Flash-MOPD a4-x (GGUF)

A mixed low-bit GGUF of XiaomiMiMo/MiMo-V2.6-Flash-MOPD (the text model), sized for a 128 GB AMD Strix Halo machine (Ryzen AI Max+ 395, Radeon 8060S). It is the MiMo-V2.6-Flash weight file that Talus ships with. It is quantized directly from Xiaomi's released weights, with an importance matrix computed on a public calibration set. Nothing in it comes from a third-party quantization.

Files

File	Size	SHA-256
MiMo-V2.6-Flash-a4-x-00001-of-00003.gguf	5.95 MB (metadata only)	b31e4530e9d0735bcd98b1f6f647fbec406594e0e4ab77134589d8fb9afd0f49
MiMo-V2.6-Flash-a4-x-00002-of-00003.gguf	48.73 GB	8c02693229bde8c7266ac2d06727ca55387024f50c0bb1eb002d044d886d31f1
MiMo-V2.6-Flash-a4-x-00003-of-00003.gguf	46.63 GB	9b0243a7901eff27006a8dcb49487d4d5984034e7332048bcbb5a48e0ad4fbe9
mimo-a4x.imatrix.gguf	992.39 MB (importance matrix)	958147270e4e3037bb291cdab64ee2949de95e2c31cdd53e7734e41ccf4f4809

Every file, including the recipe, is listed in SHA256SUMS. Point Talus or llama.cpp at the first shard; the others are found next to it.

Drafter: Talus uses the DFlash drafter dflash-MiMo-V2.6-Flash-MOPD-Q8_0.gguf from ggml-org/MiMo-V2.6-Flash-MOPD-GGUF (MIT) unchanged, so it is not re-hosted here. The 64K draft vocabulary mimo-dv-65536.bin is in j2ipper/draft-vocab.

Tensor types

508 tensors in three shards, 88.81 GiB.

Tensors	Type
Routed experts (layers 1-47)	per layer: IQ1_M / Q2_K / IQ2_XXS / IQ2_XS / IQ3_XXS / IQ3_S (most layers IQ2_XXS gate/up and IQ3_XXS down; full list in recipe/mimo-a4x-tensor-types.txt)
Attention attn_qkv, attn_output (all 51 layers)	Q4_K
Dense FFN (layer 0 and the MTP layers 48-50)	Q6_K
nextn.eh_proj, token_embd	Q8_0
output	Q6_K
Router, expert bias, norms, attention sinks	F32

By type: IQ2_XXS 35.58 GiB, IQ3_XXS 31.39 GiB, IQ2_XS 7.52 GiB, Q2_K 4.59 GiB, IQ1_M 4.38 GiB, Q4_K 2.50 GiB, Q6_K 1.09 GiB, IQ3_S 0.86 GiB, Q8_0 0.72 GiB, F32 0.19 GiB.

Recipe

All scripts and inputs are in recipe/.

1. Source: XiaomiMiMo/MiMo-V2.6-Flash-MOPD at revision b12f5573350bb1704513fcf119fca48e498b0423 (MIT). The routed experts are stored as MXFP4 (blocks of 32 with E8M0 scales); the dense weights as FP8 e4m3 with 128x128 block scales (the attention output, embeddings, LM head, router and MTP projection in BF16).

2. Staging (mq_stage.py): downloads every shard, checks its SHA-256 against the Hugging Face LFS id, and repacks the MXFP4 experts losslessly into ggml's block_mxfp4 layout (the same bit moves as llama.cpp's repack_mxfp4_blocks, checked by --selftest).

3. Reference GGUF (mq_convert.py): llama.cpp's convert_hf_to_gguf.py (MimoV2Model, llama.cpp bc449bc24) with the staged experts. The result is lossless: MXFP4 experts and BF16 tensors exact, FP8 dense weights dequantized to BF16.

4. Calibration text (mq_calib.py): calib-mimoq-v1.txt, about 177K tokens, from public datasets pinned by revision in calib-mimoq-v1.manifest.json:

   Source	License
   Salesforce/wikitext wikitext-2-raw train	CC-BY-SA-3.0 / GFDL
   wikimedia/wikipedia 20231101.zh (one file)	CC-BY-SA-3.0 / GFDL
   databricks/databricks-dolly-15k, in the chat template	CC-BY-SA-3.0
   openai/gsm8k train, in the chat template	MIT
   google-research-datasets/mbpp train, in the chat template	CC-BY-4.0
   HumanEval	MIT

   The calibration text itself is published under the licenses of its sources (CC-BY-SA parts share-alike).

5. Importance matrix (mq_refpass.sh): llama-imatrix on the reference GGUF (context 4096, 45 chunks, --process-output) -> mimo-a4x.imatrix.gguf.

6. Quantization (mq_quant.py): llama-quantize --allow-requantize --keep-split --imatrix mimo-a4x.imatrix.gguf --tensor-type-file mimo-a4x-tensor-types.txt --output-tensor-type q6_k --token-embedding-type q8_0 REF OUT Q4_K_M. The type file sets every quantized tensor's type (table above). --allow-requantize is needed only because the experts' source format is MXFP4, the model's original storage.

7. Metadata (mq_meta.py, shard 1): name, general.quantized_by, general.base_model.* = XiaomiMiMo/MiMo-V2.6-Flash-MOPD; the imatrix keys hold file names only. mq_verify.py compares two tensor layouts; a4x-layout.json is the layout of this file.

Use it with Talus

Build Talus (Quick start), download this repository, the drafter and mimo-dv-65536.bin, extract the tokenizer as the Quick start shows, and serve with a run config like this one (the arguments are Talus's shipped defaults for this model, from docs/MODELS.md):

{
  "exe": "rust/target/release/talus-cli",
  "args": ["--model", "/models/MiMo-V2.6-Flash-a4-x-00001-of-00003.gguf",
           "--max-context", "196608",
           "--ubatch", "2048",
           "--dflash-gguf", "/models/dflash-MiMo-V2.6-Flash-MOPD-Q8_0.gguf",
           "--spec", "4",
           "--spec-draft-vocab", "/models/mimo-dv-65536.bin"],
  "tokenizer": "/models/mimo-v2.6-flash/tokenizer",
  "model_name": "mimo-v2.6-flash",
  "stop_tokens": ["<|im_end|>", "<|endoftext|>"],
  "env": {},
  "lib_dirs": ["/path/to/rocm/lib", "/path/to/rocm/lib/llvm/lib"]
}

STRATA_API_KEY=<choose-a-key> rust/target/release/talus-server --config mimo.json --host 0.0.0.0 --port 8000

Talus serves it with 192K context. It also loads in llama.cpp builds that support the mimo2 architecture.

Quality

Measured by the Talus freeze run rel-final-e6cf (Talus commit e6cf959e1cad, 2026-10-28); methods in Talus's docs/METHODOLOGY.md:

Gate	Result
KL divergence vs the original weights	KLD 0.1721, top-1 89.1% on the gate text; wikitext-2: KLD 0.2255, top-1 82.5% (vs the original weights)
Needle in a haystack, 60K / 81K	15/15 / 15/15
Agentic coding suite	5/5

Speed on Strix Halo: Talus's docs/RESULTS.md.

License and credits

MIT, the license of the base model. The full license text is in LICENSE.

• Base model: XiaomiMiMo/MiMo-V2.6-Flash-MOPD, Copyright (c) 2026 Xiaomi Corporation, MIT License. The license is the MIT declaration on the model card; the model repository ships no LICENSE file of its own.
• Calibration data: the datasets listed above, under their own licenses.
• Quantized with llama.cpp (MIT).