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:Transformer0.0K Open Weights Featherless Exclusive Cold

Jackrong/Qwopus3.8-27B-Flash is a 27 billion parameter language model fine-tuned from Qwen3.8-27B, designed to optimize inference efficiency and reduce reasoning costs for agentic workloads. It achieves 12.8% faster decoding and 14.6 percentage points higher MTP draft acceptance compared to its base model, while maintaining strong general capabilities. This model is particularly suited for long-running agent workflows and resource-constrained environments where wall-clock time and token generation costs are critical.

Loading preview...

Qwopus3.8-27B-Flash: Optimized for Agentic Workloads

Qwopus3.8-27B-Flash is a 27 billion parameter model fine-tuned from the Qwen3.8-27B foundation. Its core philosophy is to preserve strong general capabilities while substantially lowering reasoning cost and response time for practical, long-running agent workloads. This model addresses the increased reasoning cost associated with iterative agent tasks, where a single task can involve dozens or hundreds of model calls.

Key Optimizations & Performance

  • Faster Decoding: Achieves 12.8% faster decoding across Math, Physics, and Chemistry subjects compared to the base Qwen3.8 model.
  • Efficient Agent Completion: Demonstrates 80.7% weighted MTP draft acceptance, a 14.6 percentage point increase over the base model, leading to more efficient agent completion.
  • Reduced Runaway Reasoning: Reduces aggregate raw_output characters by 9.9%, suggesting less pathological long-tail reasoning while retaining necessary depth.
  • Agentic Software-Engineering Battery: Passed 13 out of 14 (93%) tasks in a strict agentic software-engineering battery, completing in 26.0 minutes on an RTX 5090.
  • Trade-off: While optimizing for efficiency, the model shows a slight decrease in MMLU-Pro mixed-set accuracy (91.28% vs. 92.73% for the base model).

Good for:

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

Finetunes
4 models