Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: cortiq
|
| 4 |
+
base_model: MiniMaxAI/MiniMax-M3.1-Flash
|
| 5 |
+
base_model_relation: quantized
|
| 6 |
+
pipeline_tag: text-generation
|
| 7 |
+
tags: [cmf, cortiq, quantized, moe, q4tp, q2tp, vulkan, wgpu]
|
| 8 |
+
language: [en, zh]
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# MiniMax-M3.1-Flash — CMF
|
| 12 |
+
|
| 13 |
+
[MiniMax-M3.1-Flash](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash) 428B MoE (60 layers: 3 dense + 57×128 routed top-4 + 1 shared, SwiGLU-OAI, block-sparse 128/16, 6144 hidden, 64Q/4KV). Revision `54e4e69` as CMF for `cortiq` (Rust runtime, no Python needed to run). Requires `cortiq 0.8.31+`. Dense 2048-token fallback, sparse index tensors retained for future native sparse. Text-only.
|
| 14 |
+
|
| 15 |
+
## Files
|
| 16 |
+
|
| 17 |
+
| file | size | VRAM | notes |
|
| 18 |
+
|---|---|---:|---|
|
| 19 |
+
| `minimax-m3.1-flash-q4tp.cmf` | 212 GB (227 GB on disk with header) | 32 GB+ | **recommended**: routed experts Q4TiledP v2 (place+31-rung+unshrink), skeleton Q8_2f/f16 |
|
| 20 |
+
| `minimax-m3.1-flash-q2tp-mix.cmf` | ~166 GB | 24 GB+ | **small**: gate/up experts Q2TiledP ~2.17bpw (v2 encoder), down+skeleton Q4TiledP/Q8_2f — 35% smaller expert payload (coming) |
|
| 21 |
+
|
| 22 |
+
Both share the same decoded source; only the encoder differs. `q2tp-mix` uses the profile from `ENGINEERING.md#q2tp: encoder v2` (Q2TiledP=16, not a new dtype).
|
| 23 |
+
|
| 24 |
+
## Quick start
|
| 25 |
+
|
| 26 |
+
```sh
|
| 27 |
+
cargo install cortiq-cli --locked
|
| 28 |
+
hf download infosave/MiniMax-M3.1-Flash-cmf --include "minimax-m3.1-flash-q4tp.cmf" --local-dir .
|
| 29 |
+
cortiq verify minimax-m3.1-flash-q4tp.cmf
|
| 30 |
+
cortiq run minimax-m3.1-flash-q4tp.cmf --prompt "Explain Mars in two sentences." --greedy --max-tokens 128
|
| 31 |
+
# small edition (when available)
|
| 32 |
+
hf download infosave/MiniMax-M3.1-Flash-cmf --include "minimax-m3.1-flash-q2tp-mix.cmf" --local-dir .
|
| 33 |
+
cortiq run minimax-m3.1-flash-q2tp-mix.cmf --prompt "Explain Mars in two sentences." --greedy
|
| 34 |
+
```
|
| 35 |
+
|
| 36 |
+
## Source & provenance
|
| 37 |
+
|
| 38 |
+
* Base: [MiniMaxAI/MiniMax-M3.1-Flash](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash) @ `54e4e69dfc452b1e94a426c24a4fdeeaa8e3ad5e`
|
| 39 |
+
* Tool: `cortiq 0.8.31` (`cargo install cortiq-cli --locked`)
|
| 40 |
+
* Verify: `cortiq verify FILE` and `cortiq info FILE` print header-recorded source commit and tensor inventory (22905 tensors, 426.19B params).
|
| 41 |
+
|
| 42 |
+
## Quantization notes
|
| 43 |
+
|
| 44 |
+
* Native MXFP8/NVFP4 source decoded to f32 then quantized — no direct passthrough.
|
| 45 |
+
* `q4tp` = Q4TiledP (group 32, f16 scale), `q2tp-mix` = gate/up (`*.experts.*.gate_proj/up_proj`, `*.shared_expert.*`) on Q2TiledP, down+attention+embed on Q4TiledP/Q8_2f/f16, via `profile_quant`.
|
| 46 |
+
* Encoder v2 is default (`CMF_Q2TP_ENCODER=v1` restores old absmax-pinned).
|
| 47 |
+
|