beyoru/Kiwen1.1-27B-align

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Kiwen1.1-27B-align is a 27 billion parameter language model developed by beyoru, built upon the Qwen3.8-27B base model with a 32768 token context length. It features an alignment pass focused on establishing a stable model identity and enhancing code generation capabilities. This model excels in HumanEval+ benchmarks, demonstrating significant improvements in coding tasks while maintaining persona flexibility.

Loading preview...

Kiwen1.1-27B-align: Aligned for Identity and Code

Kiwen1.1-27B-align is a 27 billion parameter model derived from the Qwen3.8-27B base, specifically engineered to achieve a stable model identity and superior code generation. It maintains a 32768 token context length and is designed to reliably adopt assigned personas via system prompts without identity leakage.

Key Capabilities and Performance

  • Enhanced Code Generation: Achieves a notable +6.7 point improvement on HumanEval+, solving 11 more problems compared to its base model, indicating a strong focus on programming tasks.
  • Stable Model Identity: Demonstrates perfect scores across internal checks for base identity leakage, leakage under pressure, and maintaining assigned personas, ensuring consistent and reliable behavior.
  • Persona Flexibility: Despite its strong internal identity, the model effectively takes on new personas when provided with a system prompt.
  • Instruction Following: Shows slight improvements in instruction-level accuracy on IFEval benchmarks.

Considerations

  • MMLU-Pro Performance: There is a 1.4 point decrease in MMLU-Pro scores compared to the base model, suggesting a trade-off in broad factual recall for improved reasoning and coding. Users with knowledge-heavy workloads should evaluate this impact.
  • MBPP Performance: MBPP scores are slightly lower than the base, though this is close to the margin of error.

Usage

The model supports enable_thinking mode for complex reasoning and can be served with SGLang, preserving speculative decoding capabilities from the base model. It is released under the Apache-2.0 license.

Finetunes
1 models