MiniMax-M3.1-Flash-GGUF-metrics

# MiniMax M3.1 Flash GGUF: metrics and build data Everything behind [AtomicChat/MiniMax-M3.1-Flash-GGUF](https://huggingface.co/AtomicChat/MiniMax-M3.1-Flash-GGUF): the importance matrix and calibration text used for the Atomic Dynamic (AD) files, the quantization logs, and the quality measurements. The quality table itself is in the model card. | Path | What it holds | |---|---| | `imatrix/imatrix.gguf` | The importance matrix: 2,375,680 tokens through a Q8_0 copy of the model. Every one of the 128 experts in all 57 MoE blocks received tokens. | | `imatrix/imatrix.stats.txt` | Per tensor: matrices, matrices that received no tokens, mean activation energy per token, token counts. | | `calib/minimax-m3.1-moe/` | The calibration build: about 4.75M tokens of agentic, code, reasoning, 30-language and structured text rendered through M3.1 Flash's own chat template, plus 759k tokens of long documents; the recipe and the build logs. | | `logs/` | One quantization log per AD file, the hot/cold block bands chosen from the imatrix, the run log. | | `metrics/` | KLD and top-1 agreement against Q8_0 for every file, `results.json`, the raw `llama-perplexity` logs. | | `diag/` | The perplexity checks of 23 October 2026: plain text against the model's chat format, CPU against GPU, ARM weight repacking. | ## How the importance matrix was made - llama.cpp `f5027fb21307`, `llama-imatrix` on a Q8_0 copy: the BF16 (794 GiB) does not fit the GH200 we used (95 GB HBM, 525 GiB RAM). The experts of 53 of 57 MoE blocks sat in pinned host memory (`--load-mode none`, `--no-repack`). - Input: every other 512-token piece of the calibration build, 64 sequences per batch, special tokens parsed. - One local change to `llama-imatrix`: it asked for the logits of every 4th token only, because the logits of 32,768 tokens over a 200,064-token vocabulary overflow 32-bit indexing. The last block's experts therefore saw a quarter of the tokens: between 293 and 66,450 per expert. - Bands: in every AD file the blocks with the highest activation energy on `ffn_gate_exps` (the last six to eight) get one step more precision and the coldest early blocks one step less (`logs/bands-*.log`). ## Reading the measurements - **Measured on chat-format text.** KLD and top-1 are taken against Q8_0 on the odd pieces of the calibration build (the imatrix used the even ones), in M3.1's chat markup, 4096-token context. - **Absolute levels are high for this model.** Even the 4-bit files pick Q8_0's top token about three times in four. Compare the files with each other, not with other models' numbers; we are looking into why. - **Plain-text perplexity is not a quality measure here.** On wikitext the 2-bit file scores 52 against BF16's 156 (`diag/`): a chat-only model is confidently wrong on plain prose, and quantization noise makes it less so. - **On ARM CPUs** llama.cpp's weight repacking broke this MoE's Q8_0 (perplexity near 5,000). Every number here was taken with `--no-repack`. - `logs/*AD-IQ2_S*` belong to a file we did not publish: its `iq2_s` experts answered with nonsense on every backend we tried, and AD-Q2_K replaced it. ## License The importance matrix and the logs derive from MiniMax M3.1 Flash and fall under the [MiniMax Community License](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/LICENSE). The calibration text comes from [AtomicChat/calib-corpora](https://huggingface.co/datasets/AtomicChat/calib-corpora), whose sources and licenses are listed there. Made by [AtomicChat](https://atomic.chat).