Qwen3.8-Flash-Next-0.2B-A0.2B-MTP

Tiny version of Qwen/Qwen3.8-Flash-Next with MTP (Multi-Token Prediction) layers included, intended for testing MTP quantization pipelines in llm-compressor.

Base tiny model: inference-optimization/Qwen3.8-Flash-Next-0.2B-A0.2B

What's different from the base tiny model

• model_mtp.safetensors is added with 22 synthetic MTP tensors under model.language_model.mtp.layers.0.*
• model.safetensors.index.json is updated to reference both weight files
• MTP tensors are zero-initialized bfloat16, scaled to match the tiny model's dimensions (hidden_size=256, num_experts=8)
• MTP layer type: full_attention (per the original model's mtp.layer_types config)

Config

Field	Value
model_type	qwen4_exp
hidden_size	256
num_hidden_layers	4
mtp_num_hidden_layers	1
num_experts	8

Notes

This model is not intended for inference. It is a synthetic test fixture for validating MTP quantization code paths in llm-compressor. The MTP tensor names and shapes are derived from the tiny model's full-attention layer (layer 3) structure.

The base tiny model omits MTP weights per the tiny-model convention. This variant restores them for pipeline testing purposes.