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.

Mistral-Small-4-119B APEX GGUF

APEX (Adaptive Precision for EXpert Models) quantizations of Mistral-Small-4-119B-2603.

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

Benchmark Results

Benchmarks coming soon. For reference APEX benchmarks on the Qwen3.5-35B-A3B architecture, see mudler/Qwen3.5-35B-A3B-APEX-GGUF.

Available Files

File	Profile	Size	Best For
Mistral-Small-4-119B-APEX-I-Balanced.gguf	I-Balanced	~72 GB	Best overall quality/size ratio
Mistral-Small-4-119B-APEX-I-Quality.gguf	I-Quality	~62 GB	Highest quality with imatrix
Mistral-Small-4-119B-APEX-Quality.gguf	Quality	~62 GB	Highest quality standard
Mistral-Small-4-119B-APEX-Balanced.gguf	Balanced	~72 GB	General purpose
Mistral-Small-4-119B-APEX-I-Compact.gguf	I-Compact	~48 GB	Multi-GPU setups, best quality/size
Mistral-Small-4-119B-APEX-Compact.gguf	Compact	~48 GB	Multi-GPU setups
Mistral-Small-4-119B-APEX-I-Mini.gguf	I-Mini	~38 GB	Smallest viable
mmproj.gguf	Vision projector	~827 MB	Required for image understanding

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: Mistral-Small-4-119B-2603 (Mistral4/DeepSeek-V2 style)
Layers: 36
Experts: 128 routed + 1 shared (4 active per token)
Total Parameters: ~119B
Active Parameters: ~11-12B per token
Attention: Multi-head Latent Attention (MLA, kv_lora_rank=256, q_lora_rank=1024)
Vision: Pixtral encoder (mmproj included)
Context: 1M tokens (YaRN RoPE)
APEX Config: 5+5 symmetric edge gradient across 36 layers, MLA-aware tensor mapping

Run with LocalAI

local-ai run mudler/Mistral-Small-4-119B-2603-APEX-GGUF@Mistral-Small-4-119B-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.