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.

Gemopus 4 26B-A4B APEX GGUF

APEX (Adaptive Precision for EXpert Models) quantizations of Jackrong/Gemopus-4-26B-A4B-it-Preview.

Brought to you by the LocalAI team | APEX Project

Available Files

File	Profile	Size	Best For
gemopus-4-26B-A4B-APEX-I-Quality.gguf	I-Quality	20 GB	Highest quality with imatrix
gemopus-4-26B-A4B-APEX-Quality.gguf	Quality	20 GB	Highest quality standard
gemopus-4-26B-A4B-APEX-I-Balanced.gguf	I-Balanced	19 GB	Best overall quality/size ratio
gemopus-4-26B-A4B-APEX-Balanced.gguf	Balanced	19 GB	General purpose
gemopus-4-26B-A4B-APEX-I-Compact.gguf	I-Compact	15 GB	Consumer GPUs, best quality/size
gemopus-4-26B-A4B-APEX-Compact.gguf	Compact	15 GB	Consumer GPUs
gemopus-4-26B-A4B-APEX-I-Mini.gguf	I-Mini	13 GB	Smallest viable, fastest inference
gemopus-4-26B-A4B-F16.gguf	F16	48 GB	Full precision reference

Benchmark Results (Native Evals)

Model	Size	PPL	KL mean	HellaSwag	Winogrande	MMLU	ARC	TruthfulQA	pp512 t/s	tg128 t/s
APEX-I-Quality	19G	1223.5	0.532	50.5	59.2	32.1	35.1	31.0	5632	145.9
APEX-Quality	19G	1203.1	0.579	49.0	58.5	33.7	36.8	29.3	5623	143.5
APEX-I-Balanced	18G	1216.4	0.600	50.0	57.2	32.6	33.4	29.9	6211	149.4
APEX-Balanced	18G	1117.9	0.702	47.8	57.2	33.6	34.1	31.1	6221	145.7
APEX-I-Compact	14G	1258.5	0.943	49.0	59.0	32.6	34.1	30.1	6612	146.7
APEX-Compact	14G	782.1	1.617	48.8	58.2	33.5	34.4	30.0	6517	142.2
APEX-I-Mini	12G	1915.3	1.907	52.0	58.2	34.4	33.4	30.8	5904	146.8
F16 (ref)	48G	1215.9	-	-	-	-	-	-	2718	97.9

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.

Architecture

Base Model: Jackrong/Gemopus-4-26B-A4B-it-Preview
Architecture: Gemma 4 26B-A4B (MoE)
Layers: 30
Experts: 128 routed (8 active per token)
Total Parameters: 26B
Active Parameters: ~4B per token
APEX Config: 5+5 symmetric edge gradient across 30 layers
Calibration: v1.2 diverse dataset

Run with LocalAI

local-ai run mudler/Gemopus-4-26B-A4B-it-Preview-APEX-GGUF@gemopus-4-26B-A4B-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.