Jackrong/Qwopus3.8-27B-Flash

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 3, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Featherless Exclusive Cold

Qwopus3.8-27B-Flash is a 27 billion parameter model fine-tuned from Qwen3.8-27B by Jackrong, designed for efficient agent workloads. It prioritizes reduced reasoning cost and faster response times, achieving 12.8% faster decoding and 14.6 percentage points higher MTP draft acceptance compared to its base. This model is optimized for long-running agent tasks where inference efficiency and consistent completion are critical, making it suitable for resource-constrained environments.

Loading preview...

Qwopus3.8-27B-Flash: Efficient Agent Workloads

Qwopus3.8-27B-Flash is a 27 billion parameter model fine-tuned from Qwen3.8-27B, developed by Jackrong. Its core philosophy is to maintain strong general capabilities while significantly reducing reasoning costs and response times, particularly for demanding, long-running agent tasks. The model aims to minimize ineffective computation and achieve quicker, more consistent task completion.

Key Differentiators & Performance

  • Enhanced Efficiency: Achieves 12.8% faster decoding (9.347 tok/s vs. 8.284 tok/s) and 14.6 percentage points higher MTP draft acceptance (80.7% vs. 66.1%) compared to the Qwen3.8 base model.
  • Optimized for Agents: Designed to reduce wall-clock time and generated token costs in iterative agent workflows, where models are called repeatedly.
  • Reasoning Efficiency: Demonstrates a 9.9% reduction in aggregate raw output characters, suggesting less pathological long-tail reasoning while retaining necessary depth.
  • Agentic Software Engineering: Achieved 13 out of 14 passes (93%) on a strict agentic software-engineering battery, completing tasks like multi-bug repair, runtime debugging, and performance optimization.
  • Fine-Tuning: Utilizes a two-stage process: initial SFT with high-quality filtered data (1.5 million examples reduced to 10%) followed by NVIDIA NeMo-RL + GSPO for reasoning reinforcement.

Trade-offs and Limitations

  • Accuracy vs. Efficiency: The model exhibits a slight decrease in MMLU-Pro mixed-question accuracy (91.28%) compared to the base Qwen3.8 (92.73%), reflecting its focus on inference efficiency.
  • Known Issue: May occasionally generate Python code with incorrect indentation in specific tasks.

Recommended Use Cases

  • Local MTP speculative decoding where decoder throughput and draft acceptance are critical.
  • Long-running agent workflows involving repeated tool calls and iterative edits.
  • Resource-conscious workloads where wall-clock time and generated-token cost are primary operational constraints.

Finetunes
7 models