Wen 1.1

Experimental text/code checkpoint with a proof-of-concept BranchWrite LoRA merged into Wen's weights. Load Wen-1.1-BF16.gguf without an additional adapter.

This is a training-pipeline milestone. The small pilot did not demonstrate improved BranchWrite capability: the base and adapted models passed the same 15 of 19 checks, with the same four failures.

Model file

• File: Wen-1.1-BF16.gguf
• Size: 16,244,817,856 bytes
• SHA-256: 49425fd802b04b400a1725e8afe527c405bb15e828d0275ff295e38396839f34
• Precision: mostly BF16; the three merged matrices remain F32 to preserve small adapter updates. Other tensors retain their original BF16/F32 representation.
• Architecture metadata: wen1. Use a runtime with Wen support, such as the updated Private_Context_COW.

Pilot and merge verification

The pilot used a rank-8, alpha-8 LoRA on the final layer's MLP gate, up and down projections, with nine mixed training examples and four optimizer steps. Held-out and preservation checks were evaluated separately.

The merged checkpoint was compared against the original checkpoint loaded with the adapter in native COW. All 132 tested next-token choices matched, including the 128-token greedy continuation after a 1,024-token prefix. The largest measured absolute logit difference was approximately 0.0000062. This verifies the merge on those probes; it does not establish a task-quality improvement.

The other 424 tensors were preserved byte for byte. The original checkpoint and adapter remain available locally; the old model file has been removed from the current repository tree.