Jackrong's Qwopus V2 Slashes AI Agent Reasoning Time on Local Hardware
A community fine-tune of Qwen3.8-27B trades a sliver of raw accuracy for shorter reasoning traces, faster decoding, and cleaner Python output in long agent loops.
- Qwopus3.8-27B-Flash-V2 is an Apache-2.0 post-trained fine-tune of Qwen3.8-27B optimized for agent loops.
- MMLU-Pro accuracy of 92.33% versus 92.73% for the base model, with shorter reasoning traces.
- Passed 14 of 14 tasks on a hidden agentic software-engineering battery via headless Claude Code.
- Ships in GGUF from 2-bit (10.9 GB) to BF16 (54.7 GB) with bundled MTP speculative-decoding head.
- Recommended coding config: reasoning_effort xhigh with temperature=0.3 to avoid Python indentation issues.
- In one case study, finished a workload in 640s versus 1,390s for the base, which crashed.
Community researcher Jackrong has released Qwopus V2, a post-trained variant of Qwen3.8-27B designed to shorten reasoning traces, reduce agent latency, and limit Python indentation failures. Those changes target a recurring problem in multi-step agents: small delays and token costs accumulate across every model call.
What ships
| Item | Details |
|---|---|
| License | Apache 2.0 |
| Weight formats | GGUF quantizations from 2-bit at 10.9 GB through BF16 at 54.7 GB |
| Target workloads | Local coding agents, tool-calling loops, and speculative decoding |
| Runtimes listed | llama.cpp, vLLM, SGLang, and Docker Model Runner |
| Downloads | Nearly 400,000 during the first month, according to the Hugging Face counter |
The GGUF files primarily target llama.cpp-compatible engines. Other runtimes may require a different weight format or conversion path. Hugging Face download counts measure file requests and do not represent unique installations.
Agent latency compounds by the turn
A conventional chat may invoke a model once, while an agent repeatedly reads context, reasons, calls tools, inspects results, edits files, and runs tests. An additional five seconds of reasoning across 50 turns adds 250 seconds before accounting for tool execution, network calls, or retries.
V2 adds another reinforcement-learning pass to the earlier Flash pipeline. Its stated objectives are shorter reasoning traces, fewer runaway completions, and more consistent output formatting. That profile can improve completed tasks per hour even when benchmark accuracy changes only slightly.
A heavily filtered training stack
The inherited training pipeline has three stages:
- Supervised fine-tuning: Approximately 1.5 million teacher-generated examples were cleaned and scored, with the highest-ranked 10% retained.
- Reasoning reinforcement learning: The first Flash model used NVIDIA NeMo-RL with GSPO, a group-based policy-optimization method.
- V2 reinforcement learning: The new release applies revised reward functions and training methods. The author has not disclosed the complete reward design.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
