Qwen3.8-27B

Headlines · 4

1. 8h
   Jailbroken PS5 runs Qwen3.8-27B locally at 20.9 tokens/s with GPU acceleration

   Developer abhi245y has built PS5LM Plus, a native GPU application that lets jailbroken PS5 consoles run large language models locally, building on the PS5LM llama.cpp port by AI engineer Mert Cobanov. By porting Mesa's open-source AMD RADV driver and using the Vulkan compute interface to run llama.cpp as a native app, the project raised usable memory from about 6GB to 11.4GB, enough for a heavily quantized Qwen3.8-27B that generates at roughly 20.9 tokens per second, with GPU utilization at 100% and CPU at 2% during testing. The project also includes a web chat interface and an API on LAN port 8081 that is compatible with OpenAI and Anthropic interfaces, so tools such as Claude Code and Codex can call the console-hosted model; the 16GB unified memory and the risk of console or PSN bans from jailbreaking remain the main constraints.

   IT之家 AI · 🔥 18
2. 3d
   Terminal-RSI: recursive self-improving terminal agents

   An arXiv preprint introduces Terminal-RSI, which alternates Recursive Self-Rewrite (RSR) of verified trajectories with critic-inheriting Recursive Policy-Refine (RPR), and adds Turn-Boundary Skip-Observation GAE, Family Dense Reward and Success-Prioritized Reward alongside a warm-up critic model, recursive critic handoff and stage-specific learning rates. The authors report that a five-stage run from a Qwen3.8-27B base (RSR-1, RPR-1, RSR-2, RPR-2) yields Terminal-RSI-27B, scoring 19.84% on Terminal-Bench 4.0, 86.52% on Terminal-Bench 2.1, 16.95% on Terminal-Bench 3.0 and 79.00% on Terminal-Bench Hard. On LHTB they report process reward rising from 0.21 to 0.37 and strict completion from 0/46 to 5/46, while SWR100 success doubles from 3.00% to 6.00%; these figures are the paper's own claims and have not been independently verified.

   Hugging Face · Papers · 🔥 0
3. 5d
   AirLLM claims 2.8T-parameter Kimi K3 runs on a single 4GB GPU, adds low-VRAM training

   AirLLM is an open-source inference and training library that streams weights one layer at a time, keeping only a single layer resident on the GPU to cut VRAM use sharply. According to its repository, it runs 70B models on a single 4GB card without quantization, distillation or pruning, and reports running Kimi K3 (2.8T) in 3.72GB, Qwen3.8-Flash-Next (125B) in about 5.95GB and DeepSeek-V3 (671B) in roughly 12GB. Recent updates also add small-VRAM training, fine-tuning Qwen3.8-Flash-Next (125B) under 6GB and Qwen3.8-27B at about 2GB. All VRAM and training figures are the project's own reported measurements and have not been independently verified.

   GitHub Trending（每日） · 🔥 0
4. Oct 10
   Underdog AI releases Saluki 27B, a 2-bit quantized Qwen3.8-27B

   DoNews reports that Underdog AI released a model called Saluki 27B, describing it as a 2-bit quantized version of Qwen3.8-27B. The report gives no benchmarks, license terms or release channel.

   DoNews · 🔥 0

Experience and discussion from the community

Share my Qwen3.8-27B experienceAsk about Qwen3.8-27B