EigenLabs/MiMo-V2.6-Flash-MOPD-MLX-4bit-mtp

MiMo V2.6 Flash (MOPD) — MLX 4-bit with embedded MTP

An MLX-quantized version of MiMo-V2.6-Flash-MOPD for compatible Apple silicon runtimes, with an embedded multi-token prediction (MTP) head. MOPD is Xiaomi's upgrade of the MiMo-V2.6-Flash-RL checkpoint that mitigates tool-call repetition. The text model, vision tower, audio encoder and MTP weights are included in one indexed checkpoint. No separate draft-model download is required.

Artifact

Field	Value
Base model	XiaomiMiMo/MiMo-V2.6-Flash-MOPD
Architecture	mimo_v2
Inputs	Text; vision tower, audio encoder and audio tokenizer included
Output	Text, including reasoning and structured tool calls in compatible runtimes
Quantization	MLX mixed: routed experts MXFP4/group-32, attention, dense MLP, embeddings and output projection 8-bit affine/group-64, remaining quantized modules 4-bit affine/group-64
Safetensor files, including headers	170,974,651,261 bytes, approximately 159.23 GiB, across 37 shards
Indexed arrays	1,732, including 42 MTP entries, 364 vision entries and 223 audio entries
Routed experts	141 stacked expert projections; MXFP4 values repacked from the upstream MXFP4 experts without requantization
License	MIT

The routed experts keep the upstream MXFP4 values bit for bit. The upstream FP8 attention and dense weights are dequantized and stored in MLX affine formats. The per-module precision choices, scales and offsets are recorded in config.json; the 4-bit label does not mean every array uses four bits. Disk size is not a minimum-RAM specification: runtime buffers, context length, batching and offload policy affect memory requirements.

Multimodal support

The vision tower, audio encoder, audio tokenizer, tokenizer, chat template and processor configuration are included, and the vision and audio tensors equal the upstream checkpoint's. Image and audio inputs were not evaluated for this artifact. Use a runtime that explicitly supports mimo_v2 and its multimodal preprocessing.

Embedded MTP

The checkpoint includes the three trained MTP prediction layers alongside the target model. They are stored in BF16 as language_model.model.mtp.*. They were derived from the upstream FP8 model_mtp.safetensors: the fused query/key/value projection was split with the official TP=4 layout and the FP8 blocks were dequantized. An MTP-aware runtime can propose tokens with these layers and verify them against the target model before accepting output. Speculative depth is a runtime setting, not the number of trained heads or a change to the weights.

Embedding the heads does not enable speculation by itself. Check the runtime's proposal and acceptance diagnostics. In the assessed oMLX implementation, single-request text generation used MTP with adaptive depth up to 2. Acceptance was 67–96% across agent-style multi-turn chats, and the sequence recovered correctly when MTP paused and resumed mid-request.

Download and runtime compatibility

Download the complete repository with huggingface_hub:

from huggingface_hub import snapshot_download

model_path = snapshot_download(
    repo_id="EigenLabs/MiMo-V2.6-Flash-MOPD-MLX-4bit-mtp",
    local_dir="MiMo-V2.6-Flash-MOPD-MLX-4bit-mtp",
)

For reproducible deployments, supply a full immutable commit as revision. Load the resulting directory through a runtime with native mimo_v2 support. Generic MLX model loading does not establish architecture, multimodal or MTP compatibility. Darkbloom integration has not been assessed for this artifact.

Validation and limitations

• Upstream integrity: all 128 upstream weight shards were verified against their published SHA-256 before conversion.
• Inventory and layout: the tensor inventory, shapes and dtypes match an established MLX 4-bit conversion of the same architecture, plus the embedded MTP entries.
• Unchanged components: the vision and audio tensors equal the upstream checkpoint's.
• Embedded MTP: the embedded heads load bitwise identical to the upstream MTP heads.
• Full-package integrity: every file's SHA-256 is recorded in SHA256SUMS.json.

Local oMLX checks on a Mac Studio M5 Ultra (80-core GPU, 256 GB) covered multi-turn chat, code generation, and MTP pause/resume. Measured single-request throughput:

Prompt	Prefill	Decode (MTP)
1k	about 1.9k tok/s	about 88 tok/s
8k	about 3.3k tok/s	102–110 tok/s
16k	about 3.2k tok/s	about 96 tok/s

These results do not establish BF16 equivalence, the tool-call repetition improvement on any particular workload, multimodal quality, every cache or concurrency configuration, all hardware tiers, signed/trusted deployment or hosted certification. Validate the intended workload and runtime before production use.

Attribution

Original model and training: Xiaomi MiMo. EigenLabs packaging retains the supplied model license. This is not an official Xiaomi release.

See the original model card for the upstream model's intended use and limitations, and artifact-provenance.json and SHA256SUMS.json for package provenance and file hashes.