MiniMax-SLURPY-DQ-MLX

MiniMax-SLURPY-DQ-MLX

Per-tensor mixed-precision quantization of MiniMax-SLURPY for Apple Silicon — 2.54 BPW with 498 per-tensor-projection allocations (plus 16,122 per-expert PRISM decisions collapsed into MLX's SwitchGLU format).

The full SLURPY model (228.7B params) compressed from 215 GB → 68 GB (68% reduction) using PRISM Dynamic Quantization — a per-tensor-class mixed-precision allocation derived entirely from weight structure sensitivity analysis. Zero calibration data, zero training, zero datasets.

Created by Ex0bit

---

<div align="center">

💡 Support our Research & Development efforts. PRISM Members Receive access to the latest PRISM-PRO Model drops on Day-0

</div>

---

Model Details

What SLURPY inherits

A mathematically unique Designer Baby of MiniMax-M2.5 and MiniMax-M2.7 — neither parent, entirely its own model.

SLURPY inherits M2.5's architect-first coding style and MIT freedom, absorbs M2.7's RL-tuned precision on multi-agent collaboration and real-world engineering — without a single training step.

See Ex0bit/MiniMax-SLURPY for full benchmark details.

---

PRISM Dynamic Quantization

This model uses PRISM Dynamic Quantization — a per-tensor mixed-precision allocation that assigns different quantization types to different tensor classes based on weight structure sensitivity analysis.

Unlike uniform quantization (Q3, Q4, Q5), PRISM-DQ analyzes each tensor's sensitivity to quantization error and allocates precision where it matters most. Critical tensors (attention projections, key MoE experts, lm_head) receive higher precision while less impactful tensors get aggressive compression.

PRISM produced 16,122 per-expert decisions (256 experts × 62 layers × 3 projections, plus attention and embeddings). MLX's SwitchGLU packs all 256 experts per layer-projection into a single 3D tensor sharing one bit width, so the per-expert decisions collapse to the modal bit width for each of the 186 MoE projections. The remaining 312 per-tensor decisions (attention, embeddings, lm_head, routers) retain full PRISM granularity, giving 498 effective overrides.

The model's config.json contains the per-tensor quantization overrides that mlx-lm loads natively — no custom runtime required. Apple Silicon's compiled Metal kernels automatically handle mixed-precision tensors in a single forward pass at full GPU speed.

No calibration data, no importance matrices, no training data required.

---

Architecture

Identical to MiniMax-M2.5 / M2.7 — quantization-only:

• Model type: minimax_m2 / MiniMaxM2ForCausalLM
• Parameters: 228.7B total, ~10B active (MoE)
• Layers: 62
• Hidden size: 3072
• MoE: 256 experts, top-8, sigmoid routing + learned bias
• Attention: 48 query / 8 KV heads (GQA 6:1), head_dim=128
• Quantization: MLX affine, mixed 2-6 bit
• Vocab: 200,064 tokens
• Context: up to 196,608 tokens
• Thinking: Interleaved <think>...</think> (always-on)
• trust_remote_code=True required

---

Usage on Apple Silicon

mlx-lm (CLI)

pip install mlx-lm

# Interactive chat
mlx_lm.chat --model Ex0bit/MiniMax-SLURPY-PRISM-3BPW-MLX \
  --temperature 1.0 --top-p 0.95 --max-tokens 4096

# Single prompt
python -m mlx_lm.generate \
  --model Ex0bit/MiniMax-SLURPY-PRISM-3BPW-MLX \
  --prompt "Write a Python function that reverses a linked list." \
  --max-tokens 2048 \
  --temp 1.0 --top-p 0.95

Python API

from mlx_lm import load, generate

model, tokenizer = load("Ex0bit/MiniMax-SLURPY-PRISM-3BPW-MLX")

response = generate(
    model, tokenizer,
    prompt="Write a Python function that reverses a linked list.",
    max_tokens=2048,
    temp=1.0,
    top_p=0.95,
)
print(response)

Recommended sampling parameters

Important: preserve thinking in conversation history

MiniMax-M2 uses interleaved thinking. The model outputs <think>...</think> blocks during generation. You must pass these back verbatim in conversation history. Removing them degrades performance.

---

Tool calling

Same format as base SLURPY. Tool calls use <minimax:tool_call> / </minimax:tool_call> XML wrappers:

<minimax:tool_call>
<invoke name="get_weather">
<parameter name="city">San Francisco</parameter>
</invoke>
</minimax:tool_call>

---

Hardware requirements

• Apple Silicon Mac with unified memory
• 80 GB RAM minimum (model is 68 GB; needs headroom for KV cache)
• 128 GB RAM recommended for full context length
• M2 Ultra / M3 Max / M4 Max for best throughput

For non-Apple platforms, use the FP8 Ex0bit/MiniMax-SLURPY variant with vLLM.

---

Files

• 14 MLX safetensors shards (68 GB total)
• config.json with 498 per-tensor quantization overrides (collapsed from 16,122 PRISM decisions via SwitchGLU packing)
• chat_template.jinja — M2.7's chat template with tool calling support
• modeling_minimax_m2.py / configuration_minimax_m2.py — custom model code (inherited from base)

---

License

Modified MIT — same as MiniMax-M2.5. See LICENSE for full text.

The only modification to the standard MIT license: if the Software (or any derivative works) is used for commercial products or services with more than 100 million monthly active users or more than $30M annual recurring revenue, you must prominently display "MiniMax M2" on the user interface.

---

Credits

• Creator: Ex0bit
• Base model: Ex0bit/MiniMax-SLURPY
• Parents: MiniMaxAI/MiniMax-M2.5, MiniMaxAI/MiniMax-M2.7
• Quantization engine: PRISM-DQ by Ex0bit

---

Citation

@misc{minimax-slurpy-prism-mlx-2026,
  title={MiniMax-SLURPY-PRISM-3BPW-MLX: Per-tensor mixed-precision quantization of MiniMax-SLURPY for Apple Silicon},
  author={Ex0bit},
  year={2026},
  url={https://huggingface.co/Ex0bit/MiniMax-SLURPY-PRISM-3BPW-MLX}
}