APEX MoE Quantized Models Expand to Over 30 Types

This is an update on APEX, a mixed-precision quantization strategy for MoE (Mixture-of-Experts) models. Following the existing Qwen 3.5 model, over 30 major MoE models are now available quantized using the APEX method.

Key Features and Feedback:

• Long-context retention: APEX's I-Balanced and I-Compact methods maintain higher precision for shared experts and edge layers. As a result, they show high consistency even at 32k tokens or longer contexts, without the performance degradation typically seen with standard Q4_K quantization.
• Strong coding performance: For the Qwen 3.6 35B-A3B model, the I-Compact and I-Mini tiers show performance close to the FP16 model in real-world coding tasks.

Newly Added Key Models:

• Qwen family: Qwen 3.5 (122B, 397B, etc.), Qwen 3.6 (35B-A3B), Qwen3-Coder, and more
• Frontier-class MoE: MiniMax-M2.5/M2.7 (up to 228B), Mistral-Small 4 (119B), NVIDIA Nemotron-3-Super (120B), GLM-4.7 Flash, Step-3.5 Flash, Nemotron-3-Nano (with multimodal support), and more

With this update, the I-Mini/I-Compact tiers, which can run on consumer GPUs, are expected to see even greater use.