
Tessera-27B-B-CJK-RL-LoRA
This repo holds the PEFT LoRA adapter behind ZelligeAI/Tessera-27B-B-CJK-RL. The main card has all results, limits and usage notes. In short, it continues Tessera-27B-B-CJK with eight RL updates on math and code. On AIME and HMMT 2026 it got 108 of 126 attempts right against Tessera-27B-B-CJK's 97. On new states from 16 held-out coding-agent tasks it thinks 1.11x as much as Tessera-27B-B-CJK on a typical turn (90% CI 0.98-1.26) and 0.67x as much as the base model.
The adapter applies to Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0, with the stock tokenizer and chat template. It has rank 32 and alpha 32 and covers attention, the MLP and lm_head, 1.09 GB in total. It is the full adapter, not a difference from Tessera-27B-B-CJK, so apply it to the base model.
Tinker trains separate q, k and v LoRAs in the linear-attention layers, while the Hugging Face model has one fused in_proj_qkv module. This adapter fuses them with a stacked A and a block-diagonal B, at rank 96 on that module. The maximum relative fusion error is 5.3e-7. A direct conversion without this fusion drops those deltas in 48 of 64 layers without any error message.
We evaluated this LoRA on Tinker. We merged it into the bf16 weights of the main repo and checked the merge by logprob parity, as the main card describes. We never loaded the adapter with PEFT. Its keys follow the multimodal module layout, base_model.model.model.language_model.layers.N..., so load the base with AutoModelForImageTextToText. Check the load result for missing or unexpected keys. A mismatch means the adapter did not apply, and PEFT may not warn. The merged model in the main repo avoids this step.
# Untested.import torchfrom transformers import AutoModelForImageTextToTextfrom peft import PeftModel base = AutoModelForImageTextToText.from_pretrained( "Qwen/Qwen3.8-27B", revision="1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0", dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, "ZelligeAI/Tessera-27B-B-CJK-RL-LoRA")
Use reasoning_effort="xhigh" in agent loops. The license is Apache-2.0, the same as the base model's.