infosave / MiMo-V2.6-Flash-RL-cmf

MiMo-V2.6-Flash-RL · CMF q4tp

Native CMF weights for the Cortiq inference engine , converted from Xiaomi MiMo-V2.6-Flash-RL .

English · Русский · 中文

Model release candidate. All three weight files are uploaded and SHA-256 verified. Text and greedy MTP checks pass; strict multimodal acceptance is still incomplete. Engine: Cortiq 0.7.7 · crates.io . Older published versions do not include this MiMo integration.

Choose your edition

These are two complete file sets, not two copies of the 164-GB backbone. Full includes the same text weights plus a 1.241-GB multimodal companion . To upgrade from Text-only, download only the .mm.cmf file into the same folder. Text-only does not download or load the media towers.

Sizes are decimal GB. Download guides include exact files and checksum commands.

Why are the downloads large?

MiMo is a mixture-of-experts model with approximately 310 billion total parameters , not a 15-billion-parameter checkpoint. Only a subset (roughly 15 billion parameters) is active per token, but all experts must be stored . Even a hypothetical uniform four-bit representation of 310 billion parameters would occupy about 155 GB , before scales, metadata and higher-precision tensors. This conversion intentionally retains q8 and float tensors where needed.

These are disk download sizes, not VRAM requirements . The multimodal part adds only 1.241 GB; removing it does not shrink the shared MoE backbone.

Run

Use prebuilt Cortiq 0.7.7 binaries for your platform, or build the published CLI with a Rust toolchain:

cargo install cortiq-cli --version 0.7.7 --locked
cortiq --version

The engine release and this model's quality status are separate: the engine is published, while the Full edition still has the accuracy gates listed below.

Keep the companions beside the backbone; they are discovered automatically. GPU placement adapts to the available weight budget using resident layers and a dynamic expert cache. The measured host had ample RAM for the mapped model; low-host-RAM systems have not been qualified. Allow at least 167 GB of disk space for the full weights, plus space for the engine and download cache.

# Both editions
cortiq run MiMo-V2.6-Flash-RL-q4tp.cmf --prompt "Explain why the sky is blue."
cortiq serve MiMo-V2.6-Flash-RL-q4tp.cmf --host 127.0.0.1 --port 8080

# Full edition
cortiq run MiMo-V2.6-Flash-RL-q4tp.cmf --prompt "Read the text." --image page.png
cortiq run MiMo-V2.6-Flash-RL-q4tp.cmf --prompt "Describe the clip." --video frames --video-fps 1
cortiq run MiMo-V2.6-Flash-RL-q4tp.cmf --prompt "Transcribe the speech." --audio speech.wav

MTP uses three checkpoint draft layers and GPU speculative verification for greedy decoding . Nonzero-temperature sampling does not use this greedy MTP path. MTP can accelerate some prompts but is not universally faster.

Media support: images; WAV audio, including resampling/stereo; silent video from Y4M or a frame directory. MP4, compressed audio and interleaved video/audio are not supported. Full is for media understanding , not image/audio generation.

Performance

RTX PRO 6000 Blackwell 96 GB , Vulkan, one stream:

cortiq bench MiMo-V2.6-Flash-RL-q4tp.cmf --core --tokens 128 --ignore-eos --json

Default-setting runs: 32.20 / 41.40 / 40.85 tok/s · median 40.85 tok/s . This is a warmed synthetic core benchmark, not a guarantee for every prompt. Natural 128-token prompts: warm plain 29.90–36.05 tok/s ; MTP 28.16–33.62 tok/s .

The ladder simulates weight budgets on one 96-GB card , not different physical cards. Temporary upload buffers exceeded the budgets in these baseline runs; a bounded-upload candidate is not included in 0.7.7 pending a controlled A/B. Do not interpret this table as physical 16–80 GB card qualification. See validation details .

Precision and validation

• Experts: upstream MXFP4 recoded to q4tp , not a direct BF16 quantization.
• Text attention, embedding/head and initial dense MLP: q8_2f ; small tensors retain their required float formats.
• Full companion: q8_2f vision and audio tokenizer , GPTQ-q4tp audio patch encoder. The package name does not mean every tensor is four-bit.
• Recorded Wikitext PPL: 4.478 , source 4.361 , same four 512-token windows.
• CPU/full-GPU/24-GB-budget PPL128: 3.647 . EN/RU/code: all 128 token IDs match between plain/MTP and repeated runs. Separate 456-token prompt + 64-token CPU/GPU greedy continuation: exact IDs.
• Full passes OCR, shapes, chart reading and five ASR clips ( WER 0% ). Open gates: strict quantized-vision row cosine and the precise video timestamp answer. Exact-weight default-GPU tower precision now passes all six reference fixtures (maximum relative error 3.663e-5).

These open checks prevent claiming fully qualified multimodal inference. Detailed results and limitations are kept separate from the quick start.

Verify the download

On macOS, use shasum -a 256 -c in place of sha256sum -c .

Every weight file has a matching .sha256 file. All three also passed cortiq verify : envelope, tensor directory and per-tensor hashes.

sha256sum -c MiMo-V2.6-Flash-RL-q4tp.cmf.sha256
sha256sum -c MiMo-V2.6-Flash-RL-q4tp.mtp.cmf.sha256
# Full only:
sha256sum -c MiMo-V2.6-Flash-RL-q4tp.mm.cmf.sha256
cortiq verify MiMo-V2.6-Flash-RL-q4tp.cmf

Generated example

Download the interactive aquarium HTML · Original prompt · Generation and QA notes

Browser-tested at 1280×800 and 390×844, with no captured console warnings/errors. The original run reached 16,000 tokens; a model continuation supplied the ending. The downloadable example includes documented, minimal manual QA fixes for waypoints, tail geometry and bubble timing. The unaltered assembled model output is also provided. This is not presented as a successful default one-shot result.

---

Русский

Две комплектации q4tp:

• Только текст — 164.851 ГБ : основа + MTP.
• Полная — 166.093 ГБ : та же основа + MTP + MM для изображений, видео и аудио. Все файлы уже загружены.

Для перехода к полной достаточно добавить 1.241 ГБ .mm.cmf рядом с основой. Повторно скачивать 164 ГБ не нужно. Текстовые запросы не загружают медиабашни.

cortiq run MiMo-V2.6-Flash-RL-q4tp.cmf --prompt "Почему небо голубое?"

Нужен Cortiq 0.7.7 ; установка из crates.io: cargo install cortiq-cli --version 0.7.7 --locked . Размещение на GPU и обнаружение компаньонов автоматические. GPU-MTP работает при жадном декодировании; для обычной генерации с ненулевой температурой этот путь не используется.

На RTX PRO 6000 96 ГБ медиана core-бенчмарка — 40.85 ток/с ; при бюджете 64000 MiB — 44.48 ток/с . Таблица выше — бюджеты на одной карте, не тесты физических карт всех размеров. Скорость на естественных промптах отличается.

Большой размер — это хранение всех ~310 млрд параметров MoE, а не только активных на токен ~15 млрд. Это объём на диске, не требование к видеопамяти.

Веса остаются кандидатом с незакрытыми проверками Full , отдельно от выпущенного движка 0.7.7. Текстовые проверки проходят; OCR, фигуры, диаграмма и пять ASR-клипов проходят. Строгие проверки vision и временной отметки видео ещё не закрыты. Точность GPU-башни с точными весами проверена: все шесть тестов прошли. Подробности — в отчёте . Лицензия исходных весов — MIT .

---

中文

提供两种 q4tp 套件：

• 纯文本版 · 164.851 GB ： 主模型 + MTP。
• 完整多模态版 · 166.093 GB ： 相同主模型 + MTP + 图像、视频和音频输入编码器。所有权重文件均已上传。

升级只需将 1.241 GB .mm.cmf 放在主模型旁，无需重新下载 164 GB 主干。 纯文本请求不加载媒体编码器。使用已发布的 ** Cortiq 0.7.7 **； 也可通过 cargo install cortiq-cli --version 0.7.7 --locked 安装：

cortiq run MiMo-V2.6-Flash-RL-q4tp.cmf --prompt "天空为什么是蓝色的？"

GPU 放置和伴随文件发现均为自动。GPU MTP 推测验证用于贪心解码， 不用于非零温度采样，也不保证所有提示均加速。

RTX PRO 6000 96 GB 上 core 基准中位数 40.85 token/s ；64000 MiB 权重预算下为 44.48 token/s 。上表为同一显卡的预算测试，不代表已验证 所有容量的实体显卡。自然语言提示的速度有所不同。

下载需保存全部约 3100 亿 MoE 参数，而不是每个 token 激活的约 150 亿。 这是磁盘占用，不是显存要求。

引擎 0.7.7 已发布，但模型权重仍是 候选版本 ：文本、OCR、形状、图表和五段 ASR 测试通过； 严格量化视觉精度及视频时间戳验收尚未完成；默认 GPU 精确权重编码器的六项参考测试均已通过。 完整限制见 验证报告 。权重沿用 MIT 许可证。