hardin/Webcoda-AI-27B

Webcoda AI — 27B (merged full model)

Full merged weights of Webcoda AI, a knowledge assistant fine-tuned from unsloth/Qwen3.6-27B (dense, bf16) on Webcoda company Q&A. Passed 6/6 factual validation (identity, founding, leadership, partnerships, Tresillian=Umbraco, early projects) plus a no-overcorrection regression check.

Recommended: use the GGUF

For local inference on Mac / AMD / NVIDIA, use the quantized GGUF instead — it is smaller and smoke-tested: hardin/Webcoda-AI-27B-GGUF (Q6_K, ~22 GB). The LoRA adapter alone is at hardin/Webcoda-AI-27B-LoRA.

Notes on these weights

• Fine-tune: LoRA (r=128) on the language-model attention+MLP projections, merged to 16-bit. Vision-tower weights are unchanged base weights.
• MTP head removed: the base model's multi-token-prediction (nextn) speculative-decoding head was stripped. This has no effect on generated outputs — standard autoregressive inference is identical; only MTP-based speculative decoding is unavailable.
• Tokenizer: tokenizer_class set to PreTrainedTokenizerFast for portability (the base shipped a newer TokenizersBackend class).
• Architecture Qwen3_5ForConditionalGeneration requires a transformers build with qwen3_5 support.