| --- |
| base_model: MiniMaxAI/MiniMax-M2.7 |
| library_name: mlx |
| pipeline_tag: text-generation |
| license: other |
| license_name: minimax-model-license |
| license_link: https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE |
| tags: |
| - minimax |
| - m2.7 |
| - moe |
| - quantized |
| - turboquant |
| - kv-cache-quantization |
| - mlx |
| --- |
| |
| > [!NOTE] |
| > **Status (2026-07-07): no weights published yet.** This repository currently contains only the model card — it marks a planned variant that has not been released. Follow the repo to be notified when files land. |
| <!-- status-note --> |
|
|
| > [!TIP] |
| > **KV-cache quantization without any fork (recommended, 2026):** upstream |
| > llama.cpp/Ollama now cover this natively — use `-ctk q8_0 -ctv q8_0` |
| > (~half KV memory, negligible quality loss: perplexity +0.002–0.05) or |
| > `-ctk q4_0 -ctv q4_0` (~quarter memory, ≈7.6% perplexity increase). In |
| > Ollama: `OLLAMA_KV_CACHE_TYPE=q8_0` with `OLLAMA_FLASH_ATTENTION=1`. Keep |
| > K and V types symmetric to stay on the fast fused Flash-Attention path. |
| > Since April 2026, mainline llama.cpp also applies Hadamard rotation to |
| > KV activations ([PR #21038](https://github.com/ggml-org/llama.cpp/pull/21038)), |
| > which greatly improves low-bit KV quality (opt-out: |
| > `LLAMA_ATTN_ROT_DISABLE=1`). |
| > |
| > The RotorQuant/TurboQuant fork flow below is **experimental/legacy**: the |
| > TurboQuant llama.cpp PR was closed without merging (June 2026) and the fork |
| > is unmaintained relative to mainline. It is NOT required to use this model. |
| <!-- kv-upstream-note --> |
| |
| # MiniMax-M2.7-TurboQuant-MLX-2bit |
| |
| **MLX 2-bit quantized variant of [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) with TurboQuant KV-cache compression, optimized for Apple Silicon.** |
| |
| ## Overview |
| |
| MiniMax-M2.7 is a massive 256-expert Mixture-of-Experts (MoE) model with 8 experts active per token, totaling approximately 456 billion parameters. This variant combines **2-bit MLX weight quantization** with **TurboQuant** KV-cache quantization for deployment on Apple Silicon hardware. |
| |
| TurboQuant uses asymmetric per-channel quantization on the KV cache. At 2-bit, this is the most aggressively compressed variant -- it enables running on 128 GB Apple Silicon but comes with meaningful quality degradation. Best suited for experimentation, prototyping, and latency-sensitive applications where approximate outputs are acceptable. |
| |
| | Property | Value | |
| |---|---| |
| | Architecture | MoE (256 experts, 8 active/token) | |
| | Total Parameters | ~456B | |
| | Layers | 62 | |
| | Hidden Size | 3072 | |
| | Attention Heads | 48 | |
| | Weight Quantization | 2-bit (MLX) | |
| | KV-Cache Quantization | TurboQuant | |
| | Estimated Size | ~110 GB | |
| | Base Model | MiniMaxAI/MiniMax-M2.7 | |
| |
| ## Quickstart |
| |
| ```bash |
| pip install mlx-lm |
| ``` |
| |
| ```python |
| from mlx_lm import load, generate |
|
|
| model, tokenizer = load("majentik/MiniMax-M2.7-TurboQuant-MLX-2bit") |
|
|
| prompt = "What is a Comprehensive Geriatric Assessment?" |
| messages = [{"role": "user", "content": prompt}] |
| text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False) |
|
|
| response = generate( |
| model, |
| tokenizer, |
| prompt=text, |
| max_tokens=512, |
| ) |
| print(response) |
| ``` |
| |
| ## TurboQuant vs RotorQuant |
|
|
| | Feature | TurboQuant | RotorQuant | |
| |---|---|---| |
| | Technique | Asymmetric per-channel KV quantization | Rotation-based KV quantization (Hadamard transform) | |
| | Throughput | Higher throughput, lower latency | Slightly lower throughput | |
| | Quality | Good quality preservation | Better quality preservation at low bit-widths | |
| | Best For | High-throughput serving, long contexts | Quality-sensitive tasks, research | |
|
|
| > At 2-bit quantization, quality loss is significant for both methods. The [RotorQuant variant](https://huggingface.co/majentik/MiniMax-M2.7-RotorQuant-MLX-2bit) will generally produce better outputs at this bit-width due to its rotation-based outlier smoothing. |
|
|
| ## Memory Estimates (Apple Silicon) |
|
|
| | Variant | Estimated Size | Minimum Unified Memory | |
| |---|---|---| |
| | MLX 8-bit | ~456 GB | 512 GB (Mac Studio M2/M3/M4 Ultra) | |
| | MLX 5-bit | ~280 GB | 384 GB | |
| | MLX 4-bit | ~225 GB | 256 GB | |
| | MLX 3-bit | ~170 GB | 192 GB | |
| | MLX 2-bit | ~110 GB | 128 GB | |
|
|
| > **Note**: 2-bit is the most accessible variant, fitting on Apple Silicon with 128 GB+ unified memory (e.g., M2/M3/M4 Max or Ultra). Expect noticeable quality degradation compared to higher bit-widths. |
|
|
| ## See Also |
|
|
| - [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) -- Base model |
| - [majentik/MiniMax-M2.7-TurboQuant](https://huggingface.co/majentik/MiniMax-M2.7-TurboQuant) -- KV-cache only (transformers) |
| - [majentik/MiniMax-M2.7-RotorQuant-MLX-2bit](https://huggingface.co/majentik/MiniMax-M2.7-RotorQuant-MLX-2bit) -- RotorQuant MLX 2-bit |
| - [majentik/MiniMax-M2.7-TurboQuant-MLX-3bit](https://huggingface.co/majentik/MiniMax-M2.7-TurboQuant-MLX-3bit) -- MLX 3-bit |
| - [majentik/MiniMax-M2.7-TurboQuant-MLX-4bit](https://huggingface.co/majentik/MiniMax-M2.7-TurboQuant-MLX-4bit) -- MLX 4-bit |
|
|
| ## Quant trade-off (MLX lane) |
|
|
| | Bits | Approx size | Use case | Recommendation | |
| |---|---|---|---| |
| | **2-bit** | ~119 GB | Aggressive quantization | **Very low-RAM Macs** | |
| | 3-bit | ~164 GB | Lossy but small | Low-RAM Macs | |
| | 4-bit | ~192 GB | Balanced default | Recommended for most Macs | |
| | 5-bit | ~228 GB | Higher fidelity | Quality-sensitive | |
| | 6-bit | ~274 GB | Approaching FP16 quality | High-fidelity | |
| | 8-bit | ~347 GB | Near-lossless reference | Fidelity-critical work | |
|
|
| (Current variant — **2bit** — is bolded.) |
|
|
| ## Variants in this family |
|
|
| (Showing 12 sibling variants under `majentik/minimax-m2.7-*`. The current variant — `TurboQuant-MLX-2bit` — is **bolded**.) |
|
|
| | Variant | Runtime | Approx size | Use case | |
| |---|---|---|---| |
| | [RotorQuant](https://huggingface.co/majentik/minimax-m2.7-rotorquant) | runtime modifier | n/a | KV-cache root (weight-agnostic) | |
| | [RotorQuant-MLX-2bit](https://huggingface.co/majentik/minimax-m2.7-rotorquant-mlx-2bit) | mlx-lm | ~885 MB | Apple Silicon, smallest | |
| | [RotorQuant-MLX-3bit](https://huggingface.co/majentik/minimax-m2.7-rotorquant-mlx-3bit) | mlx-lm | ~1.2 GB | Apple Silicon, small | |
| | [RotorQuant-MLX-4bit](https://huggingface.co/majentik/minimax-m2.7-rotorquant-mlx-4bit) | mlx-lm | ~1.7 GB | Apple Silicon balanced | |
| | [RotorQuant-MLX-5bit](https://huggingface.co/majentik/minimax-m2.7-rotorquant-mlx-5bit) | mlx-lm | ~2.1 GB | Apple Silicon, higher fidelity | |
| | [RotorQuant-MLX-8bit](https://huggingface.co/majentik/minimax-m2.7-rotorquant-mlx-8bit) | mlx-lm | ~3.2 GB | Apple Silicon reference | |
| | [TurboQuant](https://huggingface.co/majentik/minimax-m2.7-turboquant) | runtime modifier | n/a | KV-cache root (weight-agnostic) | |
| | **TurboQuant-MLX-2bit** | mlx-lm | ~885 MB | Apple Silicon, smallest | |
| | [TurboQuant-MLX-3bit](https://huggingface.co/majentik/minimax-m2.7-turboquant-mlx-3bit) | mlx-lm | ~1.2 GB | Apple Silicon, small | |
| | [TurboQuant-MLX-4bit](https://huggingface.co/majentik/minimax-m2.7-turboquant-mlx-4bit) | mlx-lm | ~1.7 GB | Apple Silicon balanced | |
| | [TurboQuant-MLX-5bit](https://huggingface.co/majentik/minimax-m2.7-turboquant-mlx-5bit) | mlx-lm | ~2.1 GB | Apple Silicon, higher fidelity | |
| | [TurboQuant-MLX-8bit](https://huggingface.co/majentik/minimax-m2.7-turboquant-mlx-8bit) | mlx-lm | ~3.2 GB | Apple Silicon reference | |
|
|
