mudler/Huihui3.5-67B-A3B-APEX-GGUF

Huihui3.5-67B-A3B APEX GGUF

APEX (Adaptive Precision for EXpert Models) quantizations of Huihui3.5-67B-A3B.

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

Available Files

File	Profile	Size	Best For
Huihui3.5-67B-A3B-APEX-I-Balanced.gguf	I-Balanced	46 GB	Best absolute quality (with imatrix)
Huihui3.5-67B-A3B-APEX-Balanced.gguf	Balanced	46 GB	Best absolute quality
Huihui3.5-67B-A3B-APEX-I-Quality.gguf	I-Quality	41 GB	Best quality/compression ratio (with imatrix)
Huihui3.5-67B-A3B-APEX-Quality.gguf	Quality	41 GB	Best quality/compression ratio
Huihui3.5-67B-A3B-APEX-I-Compact.gguf	I-Compact	31 GB	Consumer GPUs (with imatrix)
Huihui3.5-67B-A3B-APEX-Compact.gguf	Compact	31 GB	Consumer GPUs
Huihui3.5-67B-A3B-APEX-I-Mini.gguf	I-Mini	26 GB	Smallest viable
Huihui3.5-67B-A3B-F16.gguf	F16	125 GB	Full precision source

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).

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

Architecture

• Model: Huihui3.5-67B-A3B (qwen3_5_moe)
• Layers: 40 (hybrid: linear attention + full attention every 4th layer)
• Experts: 512 routed (8 active per token)
• Total Parameters: ~67B
• Active Parameters: ~3B per token
• Origin: Expert merge of Qwen3.5-35B-A3B + Holo3-35B-A3B
• APEX Config: 5+5 symmetric edge gradient across 40 layers
• Calibration: v1.3 diverse dataset (chat, code, reasoning, multilingual, tool-calling, Wikipedia)

Run with LocalAI

local-ai run mudler/Huihui3.5-67B-A3B-APEX-GGUF@Huihui3.5-67B-A3B-APEX-I-Balanced.gguf

Credits

APEX is brought to you by the LocalAI team. Developed through human-driven, AI-assisted research. Built on llama.cpp.