hemlang/Hemlock-Qwen3.8-27B

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 19, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Hemlock-Qwen3.8-27B is a 27 billion parameter vision-language model developed by hemlang, built upon a SLERP merge of Qwen3.6-line and Qwen3.8-line base models. It is specifically fine-tuned to generate Hemlock code, achieving the highest recorded score of 33/38 (pass@1 0.816) on the HemBench suite. This model maintains general reasoning capabilities, scoring 66.22% on ARC-Challenge, and retains vision and multi-turn prediction (MTP) functionalities.

Loading preview...

Overview

hemlang/Hemlock-Qwen3.8-27B is a 27 billion parameter vision-language model primarily designed for generating Hemlock code. It is constructed by merging the Hemlock-Qwen3.8-27B-LoRA adapter into a WichtelHui-Qwen3.8-27B-SLERP base. The base model itself is a 50/50 SLERP of Wichtel-Qwen3.6-27B and Huihui-Qwen3.8-27B-abliterated, indicating a mixed lineage from Qwen3.6 and Qwen3.8 architectures.

Key Capabilities and Performance

  • Hemlock Code Generation: Achieves the highest recorded score on the HemBench suite with 33/38 tasks passed (pass@1 0.816), demonstrating strong proficiency in generating executable Hemlock code. This significantly improves upon its un-adapted Qwen3.8 parent, which often generates Rust- or C-shaped code.
  • General Reasoning: The adapter training did not degrade general abilities; the model scored 66.22% on the ARC-Challenge, an improvement over its WichtelHui base (60.54%) and Qwen3.8-27B (52.84%).
  • Vision and MTP: Retains full vision and multi-turn prediction capabilities, with all relevant tensors intact from the base model.
  • Safety: Inherits safety characteristics from its WichtelHui base, including 18/18 on ccp_truth and refusals on safety_control probes.

Training Details

The Hemlock adapter was trained on the hemlang/Hemlock-SFT-combined dataset, comprising 5,852 rows, with HemBench-overlapping rows excluded. It utilized LoRA with r=32, α=64, for 2 epochs on Qwen3.8-27B.

Usage Considerations

  • For optimal code generation, it is recommended to serve the model with thinking disabled.
  • HemBench measures code execution and output correctness, not idiomatic style or efficiency.
  • Due to its mixed lineage, behavior may not always align perfectly with pure Qwen3.8 derivatives.

Finetunes
3 models