YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Model: Qwen3.5-397B-A17B (Qwen's latest Mixture-of-Experts multimodal model) What was done:

Applied REAP (Router-weighted Expert Activation Pruning) from Cerebras Research to prune 55% of MoE experts Original model: 512 experts per layer, 10 active per token Pruned model: 230 experts per layer, 10 active per token Observation phase: 128 calibration samples from evol-codealpaca-v1 dataset, cosine similarity scoring, seed 42 Pruning method: frequency-based (experts ranked by activation frequency across calibration data, bottom 55% removed)

Key details:

Original model size: 752GB (BF16) Pruned safetensors: ~377GB (BF16) GGUF Q3_K_M: ~72GB (estimated) Architecture: 60 transformer layers, fused MoE experts (gate_up_proj + down_proj), linear attention + full attention pattern, Mamba SSM components Experts reduced from [512, 2048, 4096] โ†’ [230, 2048, 4096] per layer Router weights sliced accordingly

Tools used:

REAP: https://github.com/cerebras/reap llama.cpp for GGUF conversion and quantization

Based on research:

Cerebras REAP paper โ€” shows 50%+ expert pruning retains 95%+ baseline quality on code generation and reasoning tasks

Downloads last month
16
Safetensors
Model size
208B params
Tensor type
BF16
ยท
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support