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.

Carnice MoE 35B-A3B APEX GGUF

APEX (Adaptive Precision for EXpert Models) quantizations of samuelcardillo/Carnice-MoE-35B-A3B.

Brought to you by the LocalAI team | APEX Project

Available Files

File	Profile	Size	Best For
Carnice-MoE-35B-A3B-APEX-I-Quality.gguf	I-Quality	21 GB	Highest quality with imatrix
Carnice-MoE-35B-A3B-APEX-Quality.gguf	Quality	21 GB	Highest quality standard
Carnice-MoE-35B-A3B-APEX-I-Balanced.gguf	I-Balanced	24 GB	Best overall quality/size ratio
Carnice-MoE-35B-A3B-APEX-Balanced.gguf	Balanced	24 GB	General purpose
Carnice-MoE-35B-A3B-APEX-I-Compact.gguf	I-Compact	16 GB	Consumer GPUs, best quality/size
Carnice-MoE-35B-A3B-APEX-Compact.gguf	Compact	16 GB	Consumer GPUs
Carnice-MoE-35B-A3B-APEX-I-Mini.gguf	I-Mini	13 GB	Smallest viable, fastest inference
Carnice-MoE-35B-A3B-F16.gguf	F16	65 GB	Full precision reference

Benchmark Results (Native Evals)

Model	Size	PPL ↓	KL ↓	HellaSwag	WinoGrande	MMLU	ARC-C	TruthfulQA	pp512 t/s	tg128 t/s
F16 (ref)	65G	6.16	-	-	-	-	-	-	2315	109.1
APEX-Quality	21G	6.2	0.010	83.5	74.0	40.9	56.9	34.0	4717	134.2
APEX-I-Quality	21G	6.2	0.009	83.0	75.0	40.3	55.5	34.3	4734	132.6
APEX-Balanced	24G	6.2	0.007	83.0	73.8	41.1	54.5	33.8	4572	130.3
APEX-I-Balanced	24G	6.2	0.006	83.5	74.8	40.6	54.2	34.0	4539	128.7
APEX-Compact	16G	6.4	0.045	82.8	75.5	40.8	55.9	34.0	4516	132.1
APEX-I-Compact	16G	6.3	0.032	83.0	73.8	41.2	56.2	34.9	4352	130.6
APEX-I-Mini	13G	6.6	0.071	82.0	72.2	40.6	53.8	33.7	4293	133.1

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: samuelcardillo/Carnice-MoE-35B-A3B
Architecture: Qwen3.5-MoE 35B-A3B
Layers: 40
Experts: 256 routed (8 active per token)
Total Parameters: 35B
Active Parameters: ~3B per token
APEX Config: 6+6 symmetric edge gradient across 40 layers
Calibration: v1.2 diverse dataset

Run with LocalAI

local-ai run mudler/Carnice-MoE-35B-A3B-APEX-GGUF@Carnice-MoE-35B-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.