Qwen3.5-35B-A3B SWE SFT with packing, step 41
Final production checkpoint from run qwen35_swe_npu_packing_20261117_144822_9i032izw,
fine-tuned from Qwen/Qwen3.5-35B-A3B-Base on 1,000 SWE conversations for
3 epochs (41 optimizer steps) with online sample packing at a 128K maximum sequence length.
Weights preserve the checkpoint's FP32 parameters. Optimizer state is excluded.
The tokenizer, chat template, and generation configuration accompany the weights.
This is a language-only export with a Qwen3.5 MoE conditional-generation
configuration. Vision tower weights are not included. Use a loader that supports
language-only Qwen3.5 MoE inference, such as vLLM's language-only mode.
The export validates the checkpoint completion marker, tensor keys and shapes,
finite values, zero router observation buffers, and safetensors shard headers.
See export_manifest.json for export provenance.