benthecarman EXL3 2.27bpw (vs official FP8)
source ↗HumanEval+
88.4−1.8baseline 90.2
MBPP+
77.2+0baseline 77.2
GSM8K
95.2−1.2baseline 96.4
MMLU-Pro
73.8−2.6baseline 76.4
GPQA Diamond
50−6.2baseline 56.2
MiMo-V2.6-Flash
- Organization
- Xiaomi
- Parameters
- 309B total / 15B active
- Context
- 1M
- License
- MIT
- Release
- 2026-10
sparse MoE 48 layers (39 SWA + 9 GA), 256 experts top-8, 128-token sliding window, omni (text/image/video/audio)
MiMo-V2.6-Flash-RL is Xiaomi's 309B-total / 15B-active omni MoE with a 1M context, 128-token sliding-window attention and an MIT license. The benthecarman EXL3 build at 2.27 bpw (85.28 GiB) decodes ~31 tok/s on a DGX Spark and keeps most quality versus the official FP8, with MMLU-Pro at 76.4 → 73.8.
MiMo-V2.6-Flash is a 309B total / 15B active MIT model from Xiaomi with 1M context. Quantized checkpoints lose 0.35–6.25 points on the 7 measured per-quant benchmarks.
Base-model benchmarks
Terminal-Bench 2.1
87.6DeepSWE v1.1
67.9ProgramBench
26Toolathlon-Verified
73.6OSWorld-Verified
80.8JobBench
61.2Agents' Last Exam
27.6CyberGym
95.1
Quant checkpoints
| Checkpoint | Family | Size | Source |
|---|---|---|---|
| XiaomiMiMo official MXFP4+FP8 Official native checkpoint (MXFP4 + FP8) | MXFP4 | 177.8 GB | source ↗ |
| XiaomiMiMo MiMo-V2.6-Flash-MOPD ~140 GB MOPD-distilled checkpoint (not a quantization) | BF16 | 140 GB | source ↗ |
| ggml-org MiMo-V2.6-Flash-RL GGUF Q2_K | GGUF | 126 GB | source ↗ |
| ggml-org MiMo-V2.6-Flash-RL MXFP4 | MXFP4 | 167 GB | source ↗ |
| ProCreations MiMo-V2.6-Flash-RL-NVFP4 ~194 GB; bit-exact vs FP8, held-out PPL identical (code 2.3774, math 1.5172) | NVFP4 | 194 GB | source ↗ |
| benthecarman MiMo-V2.6-Flash-RL EXL3 2.27bpw GiB | EXL3 | 85.28 GB | source ↗ |
| pfeifferj MiMo-V2.6-Flash-RL-GSQ-RCO-GGUF GSQ-RCO 3-bit | GGUF | 115.7 GB | source ↗ |
| r0b0tlab MiMo-V2.6-Flash-RL-NVFP4 NVFP4 + FP8 KV | NVFP4 | 84.4 GB | source ↗ |
Per-quant quality deltas
Serving metrics
EXL3 decode (no drafter)
31 tok/sHardware: NVIDIA DGX Spark
Setup: benthecarman EXL3 2.27bpw
source ↗EXL3 decode (coding, DFlash drafter)
49.5 tok/s (35–80 range; reasoning 61.1)Hardware: NVIDIA DGX Spark
Setup: EXL3 2.27bpw + DFlash drafter
source ↗EXL3 decode @249K ctx (drafter)
33.6 tok/s (from 16.1 without drafter)Hardware: NVIDIA DGX Spark
Setup: EXL3 2.27bpw + DFlash drafter
source ↗
What’s not published
- The ggml Q2_K GGUF was built without an imatrix, so its quality is at risk.
- NVFP4 on 2x RTX Pro 6000 did not sustain 40–60 tok/s (CPU-offloaded).
Sources
- https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL ↗
- https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-MOPD ↗
- https://huggingface.co/ggml-org/MiMo-V2.6-Flash-RL-GGUF ↗
- https://huggingface.co/ProCreations/MiMo-V2.6-Flash-RL-NVFP4 ↗
- https://huggingface.co/benthecarman/MiMo-V2.6-Flash-RL-exl3 ↗
- https://huggingface.co/pfeifferj/MiMo-V2.6-Flash-RL-GSQ-RCO-GGUF ↗
- https://huggingface.co/r0b0tlab/MiMo-V2.6-Flash-RL-NVFP4 ↗
FAQ
- What is MiMo-V2.6-Flash-RL's license, context window and parameter count?
- MIT license, 1M context window, 309B total / 15B active parameters (Xiaomi).
- How much quality does quantization cost MiMo-V2.6-Flash-RL?
- Across 10 published deltas, quantized checkpoints differ from the full-precision baseline by 0.66 points on average (range -6.25 to 12.2).
- How many quant checkpoints are cataloged for MiMo-V2.6-Flash-RL?
- 8 checkpoints across MXFP4, BF16, GGUF, NVFP4, EXL3, sizes from 84.4 GB to 194 GB.
- When was MiMo-V2.6-Flash-RL released?
- 2026-10.
- Where do the MiMo-V2.6-Flash-RL benchmark numbers come from?
- From 7 official sources — vendor model cards, quantization publisher cards and serving recipe pages.