Price: $0.05514 -0.6417%
Market Cap: $8.65B 0.4161%
Volume (24h): 419.17M 0%
Dominance: 0.4161%
Price: $0.05514 -0.6417%
Market Cap: $8.65B 0.4161%
Volume (24h): 419.17M 0%
Dominance: 0.4161% 0.4161%
  • Price: $0.05514 -0.6417%
  • Market Cap: 8.65B 0.4161%
  • Volume (24h): 419.17M 0%
  • Dominance: 0.4161% 0.4161%
  • Price: $0.05514 -0.6417%
Home > 视频 > Someone installed LoRA on Qwen 3.8 and saved 4.5 times tokens. Are you still burning computing power?

Someone installed LoRA on Qwen 3.8 and saved 4.5 times tokens. Are you still burning computing power?

Release: 2026/11/19 01:04 Reading: 0

Original author:智用AI

Original source:https://www.youtube.com/embed/hs2BpFPKuqA

Hugging Face harness-hub's Qwen3.8-27B LoRA adapter reaches the accuracy of the basic model when thinking mode is turned off, and the token consumption is reduced by about 4.5 times. Hugging Face's harness-hub team released the LoRA adapter Qwen3.8-27B-pi-adapted for Qwen3.8-27B, specifically used for pi encoding agents. The adapter is trained through reinforcement learning methods and improves the pass@1 accuracy from 39.2% to 64.7% on SWE-bench Pro and from 39.5% to 55.8% on Terminal-Bench 2.1. More importantly, it reaches the accuracy level of the basic model with thinking mode turned off when thinking mode is turned off. At the same time, token consumption is reduced by about 4.5 times and the number of interaction rounds is significantly shortened. This article disassembles the reinforcement learning training mechanism, evaluation performance and deployment path of the LoRA adapter. Qwen3.8-27B, LoRA, coding agent, reinforcement learning, pi, SWE-bench, vLLM. If this kind of underlying disassembly is useful to you, subscribe + open the little bell, and the new update will be delivered to you as soon as possible. ⏱ Chapter 00:00 The counterintuitive result is shown first: Turn off the thinking mode, the accuracy rate rises instead of falling, and the number of interaction rounds is greatly shortened. 00:46 It's not a general fine-tuning, but a reinforcement learning adapter customized for the specific agent environment of pi. 01:14 The training scale is not large: 750 scale800 tasks plus 300 TermG 01:26 The adapter is highly bound to the pi and does not guarantee migration to other agent frameworks; at the same time, the base architecture and context window are hard 02:23 Judgment worth taking away: Doing small-scale LoRA reinforcement learning alignment for a specific agent environment is to improve local programming 🔎 Continue selection from this video: https://zhiyong.dev/y/hs2BpFPKuqA --- 🔎 Related AI tools, models and applications: https://zhiyong.dev

Selected Topics

  • Dogecoin whale activity
    Dogecoin whale activity
    Get the latest insights into Dogecoin whale activities with our comprehensive analysis. Discover trends, patterns, and the impact of these whales on the Dogecoin market. Stay informed with our expert analysis and stay ahead in your cryptocurrency journey.
  • Dogecoin Mining
    Dogecoin Mining
    Dogecoin mining is the process of adding new blocks of transactions to the Dogecoin blockchain. Miners are rewarded with new Dogecoin for their work. This topic provides articles related to Dogecoin mining, including how to mine Dogecoin, the best mining hardware and software, and the profitability of Dogecoin mining.
  • Spacex Starship Launch
    Spacex Starship Launch
    This topic provides articles related to SpaceX Starship launches, including launch dates, mission details, and launch status. Stay up to date on the latest SpaceX Starship launches with this informative and comprehensive resource.
  • King of Memes: Dogecoin
    King of Memes: Dogecoin
    This topic provides articles related to the most popular memes, including "The King of Memes: Dogecoin." Memecoin has become a dominant player in the crypto space. These digital assets are popular for a variety of reasons. They drive the most innovative aspects of blockchain.