---
license: other
license_name: minimax-license
license_link: https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE
base_model: MiniMaxAI/MiniMax-M2.7
base_model_relation: quantized
library_name: transformers
pipeline_tag: text-generation
language:
  - en
  - zh
  - ru
tags:
  - minimax
  - minimax-m2
  - moe
  - mixture-of-experts
  - bf16
  - dequantized
---

# MiniMax-M2.7 — BF16 (dequantized from FP8)

Plain `bfloat16` weights of [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7),
reconstructed from the upstream block-FP8 (E4M3, 128×128 blocks) checkpoint
via shard-by-shard blockwise dequantization. **No calibration, no rounding loss
beyond the original FP8→BF16 cast** — every block is materialized exactly:

```
bf16_block = (fp8_block.float() * scale_fp32).bfloat16()
```

## Why this exists

`MiniMaxAI/MiniMax-M2.7` ships natively in FP8. On Ampere and earlier
(e.g. RTX A5000) FP8 tensor cores don't exist and inference engines have
to emulate FP8 through FP16 — paying double bandwidth without the speed
benefit. For further offline quantization (AWQ, GPTQ, RTN INT8, …) you
need plain BF16 weights anyway: `transformers + torch_dtype=bfloat16`
won't materialize the attention projections under the FP8 quant config,
which trips up `llmcompressor`'s GPTQ tracer.

This repo is the missing intermediate: **upstream MiniMax-M2.7 weights in
plain BF16 safetensors**, ready to be fed into any standard quantization
pipeline.

## Contents

- 47 shards `model-NNNNN-of-00047.safetensors`
- rebuilt `model.safetensors.index.json` (no `*.weight_scale_inv` entries)
- `config.json` with the upstream `quantization_config` stripped
- tokenizer + custom modeling `.py` files copied verbatim from the FP8 source

Total ≈ **458 GB**.

## Provenance

Produced on a single 48 GB GPU pod (~30 minutes wall time) using a
~150-line script — see
[`dequant_fp8_blockwise.py`](https://github.com/operationrange/zonatelecom-agent/blob/main/scripts/quant/dequant_fp8_blockwise.py).

Process per shard:

1. open `model-XXXXX-of-00130.safetensors` from the FP8 source
2. for each `*.weight` (FP8 e4m3fn): look up `*.weight_scale_inv` (FP32, 128×128)
3. broadcast scale to weight shape, multiply, cast to BF16
4. drop the scale tensor
5. write `model-NNNNN-of-00047.safetensors` (5 GB shards)

Other tensors (embeddings, layer norms, MoE routers/gates that were already
unquantized in the upstream config's `modules_to_not_convert`) are passed
through with a BF16 cast.

## Quick load

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

m = AutoModelForCausalLM.from_pretrained(
    "operationrange/MiniMax-M2.7-BF16",
    torch_dtype="bfloat16",
    device_map="auto",
    trust_remote_code=True,
)
tok = AutoTokenizer.from_pretrained("operationrange/MiniMax-M2.7-BF16", trust_remote_code=True)
```

Inference at full BF16 needs ≥ ~470 GB combined GPU+CPU memory, so this
checkpoint is mostly intended as a starting point for further compression
(AWQ-INT4, GPTQ-INT8, etc.) rather than direct serving.

## License

Inherits the [MiniMax-M2 license](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE) from the
upstream model. No weights were modified — only the storage format.