modrill/code-nothink-q8b-20260908

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The modrill/code-nothink-q8b-20260908 is an 8 billion parameter language model based on Qwen3-8B-Base, fine-tuned for code generation tasks. It utilizes a LoRA adaptation method and is specifically optimized for "NoThink" code evaluation scenarios, achieving a significant performance improvement over its base model. This model is designed for developers requiring a robust code generation assistant, particularly in environments where direct code output without intermediate reasoning steps is preferred.

Loading preview...

Model Overview

modrill/code-nothink-q8b-20260908, formerly known as Qwen3-8B-Base-code-v4-nothink, is an 8 billion parameter model built upon the Qwen/Qwen3-8B-Base architecture. It has been specifically fine-tuned using a LoRA r64 / \u03b1128 method across seven projections, then merged into full-model bf16 safetensors.

Key Capabilities & Training

  • Code Generation Focus: Trained on NoThink code SFT (paired V4, physical 2-epoch concat) data, emphasizing direct code output.
  • Optimized for NoThink Evaluation: Evaluated using the official LiveCodeBench DEV256 suite in "NoThink" mode, which means it does not prefill <think> tokens.
  • Performance Improvement: Achieved a mean pass@1 score of 30.1% on DEV256, representing an approximate +19 point improvement over the base model's 22.7%.
  • Tokenizer: Uses a specific qwen3-v4-tail151643 tokenizer bundle, which is included in this repository.

Usage Notes

  • Tokenizer and Chat Template: Users should utilize the tokenizer and chat_template.jinja provided within this repository.
  • NoThink Inference: Ensure that <think> tokens are not prefilled during inference; if the template supports enable_thinking, it should be set to false.
  • Stop Tokens: Important stop token IDs are 151643 (<|endoftext|>) and 151645 (<|im_end|>).

Good for

  • Code generation tasks where direct, efficient code output is prioritized.
  • Environments requiring a specialized code model that performs well under "NoThink" evaluation paradigms.
  • Developers seeking a fine-tuned Qwen3-8B variant with enhanced coding capabilities.

Finetunes
1 models