Yanun/MiMo-V2.6-Flash-RL-MXFP4

MiMo-V2.6-Flash-RL — complete Whallm installed artifact

Complete artifact uploaded and remote integrity checked. A usable revision must contain manifest.json and every file listed in its files array, with matching sizes and SHA-256 values. Whallm must pin a verified immutable revision, not the mutable main branch.

This is a prepared installed model for Whallm on Apple Silicon, not a standard Transformers or mlx-lm checkpoint. Downloading a verified complete revision requires no local weight conversion.

Source and preservation

Source: XiaomiMiMo/MiMo-V2.6-Flash-RL, revision 5711b268169967567844e1e560e8a3966da959b1. The source model card declares MIT licensing. Original code, notices and model documentation are preserved under checkpoint/.

All 90 upstream files are retained whole or reconstructible byte-for-byte. The 64 repacked source shards passed reconstruction against their original SHA-256 values. The other 26 files were preserved whole, including:

• AudioTokenizer, original three-layer MTP, five-layer DFlash and mask embedding;
• all tokenizer, processor and model configuration files;
• original model code, model card, technical report and architecture image.

All vision and audio-patch tensors are preserved. Original common FP8 weights and scales are retained in checkpoint/common.bin; no encoder or draft weights were discarded to produce a text-only subset.

Installed layout

Format 4 contains 151 payload files / 193,730,445,975 bytes (180.426 GiB), excluding manifest.json and this model card.

• experts/: 12,032 native MXFP4 experts, 47 expert-bearing layers, 256 experts per layer, top-8 routing. Each packed expert is 13,369,344 bytes. Values were repacked, not requantized.
• common.bin: 331 prepared target-backbone tensors (48 backbone/KV layers).
• vision/common.bin: 364 prepared vision tensors.
• audio/common.bin: 95 prepared audio-patch tensors.
• mtp/common.bin: 36 prepared MTP tensors.
• checkpoint/: preserved original common payload, exact shard headers, and all companion files.
• checkpoint-map.json: original file provenance and tensor-byte reconstruction map.
• preservation.json: completed local source-preservation checks.

Common FP8 tensors are decoded to BF16. Fused QKV is decoded with the source TP=4 shard-local scale grids and reordered into canonical Q/K/V. Original FP8 bytes and scales remain available for lossless source reconstruction. This is not a claim of bitwise parity with SGLang's subsequent CUDA FP8 requantization.

Validation and limitations

The complete target backbone loaded locally and generated Hello. and 你好。. Generation resumed successfully after closing an active generation iterator. Local source preservation and all installed-file SHA-256 values were verified.

Complete data preservation does not mean complete runtime support. Vision, video and audio processing, encoder execution, MTP/DFlash decoding, document integration, and MCP remain unfinished in the development implementation. Tool/API and App end-to-end acceptance are not complete. Short text smoke tests are not full-model reference parity, quality benchmarks or speed measurements. The checkpoint's 1,048,576-token limit is not a locally validated usable context.

This model upload is not an App release. Use a Whallm build that explicitly supports mimo-v2.6-flash-rl / manifest format 4. Preserve the complete manifest file set; do not delete currently unused modalities or draft models.

Publication verification

All 151 payload files and the manifest were checked against the local artifact. LFS SHA-256 values and sizes match; Git-stored files and the manifest were downloaded anonymously and SHA-256 checked at an immutable commit. This does not claim a second full download of all 180.426 GiB or App end-to-end acceptance.