mudler/MiniMax-M2.7-APEX-GGUF

⚡ Each donation = another big MoE quantized

I host 25+ free APEX MoE quantizations as independent research. My only local hardware is an NVIDIA DGX Spark (122 GB unified memory) — enough for ~30-50B-class MoEs, but bigger ones (200B+) require rented compute on H100/H200/Blackwell, typically $20-100 per quant.
If APEX quants are useful to you, your support directly funds those bigger runs.

🎉 Patreon (Monthly)  |  ☕ Buy Me a Coffee  |  ⭐ GitHub Sponsors

💚 Big thanks to Hugging Face for generously donating additional storage — much appreciated.

MiniMax-M2.7 APEX GGUF

APEX (Adaptive Precision for EXpert Models) quantizations of MiniMax-M2.7.

Brought to you by the LocalAI team | APEX Project | Technical Report

Note: MiniMax M2 architecture support in llama.cpp is still maturing. If you encounter inference issues, ensure you're using a recent llama.cpp build and report issues upstream.

Available Files

File	Profile	Size	Best For
MiniMax-M2.7-APEX-I-Balanced.gguf	I-Balanced	155 GB	Best overall quality/size ratio
MiniMax-M2.7-APEX-Balanced.gguf	Balanced	155 GB	General purpose
MiniMax-M2.7-APEX-I-Quality.gguf	I-Quality	129 GB	Highest quality with imatrix
MiniMax-M2.7-APEX-Quality.gguf	Quality	129 GB	Highest quality standard
MiniMax-M2.7-APEX-I-Compact.gguf	I-Compact	100 GB	Multi-GPU setups, best quality/size
MiniMax-M2.7-APEX-Compact.gguf	Compact	100 GB	Multi-GPU setups
MiniMax-M2.7-APEX-I-Mini.gguf	I-Mini	80 GB	Smallest "safe" tier
MiniMax-M2.7-APEX-I-Nano.gguf	I-Nano (new)	64 GB	Experimental — IQ2_XXS mid-layer experts
MiniMax-M2.7-APEX-F16-*.gguf	F16 reference	426 GB (10 shards)	Full-precision BF16 for imatrix/further research

What is APEX?

APEX is a quantization strategy for Mixture-of-Experts (MoE) models. It classifies tensors by role (routed expert, shared expert, attention) and applies a layer-wise precision gradient — edge layers get higher precision, middle layers get more aggressive compression. I-variants use diverse imatrix calibration (chat, code, reasoning, tool-calling, agentic traces, Wikipedia).

The key insight: in MoE models, expert FFN tensors make up the bulk of model weight but only ~8/256 experts activate per token. APEX compresses middle-layer experts more aggressively while preserving edge layers (first/last 5) and keeping attention and shared-expert tensors at higher precision.

See the APEX project for full details, technical report, and scripts.

Nano (new experimental tier)

APEX M2.7 debuts the Nano tier, which pushes mid-layer routed experts to IQ2_XXS (2.06 bpw), near-edge to IQ2_S, edges to Q3_K, and keeps shared experts at Q5_K. About 20% smaller than Mini with modest quality cost, viable only on MoE thanks to sparse per-token activation. Requires imatrix.

Benchmarks for Nano are pending. Feedback welcome.

Architecture

• Model: MiniMax-M2.7 (MiniMaxM2)
• Layers: 62
• Experts: 256 routed (8 active per token)
• Total Parameters: ~228 B
• Active Parameters: ~10 B per token
• Source Format: FP8 (float8_e4m3fn, block-quantized 128×128)
• Intermediate Format: BF16 (via unsloth's pre-converted BF16 GGUF)
• APEX Config: 5+5 symmetric edge gradient across 62 layers
• Calibration: v1.3 diverse dataset (chat, code, reasoning, multilingual, tool-calling, Wikipedia)

Run with LocalAI

local-ai run mudler/MiniMax-M2.7-APEX-GGUF@MiniMax-M2.7-APEX-I-Balanced.gguf

Credits

• Base model: MiniMaxAI
• BF16 GGUF source: unsloth/MiniMax-M2.7-GGUF
• APEX quantization: LocalAI team
• Built on llama.cpp