First screen, captured 1 day ago. Hover to scroll the page, or open the video.

Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness

What HN thinks

Five small models walk into an old benchmark and claim a frontier tie; the room asks to see the contamination report.

Benchmark methodology is unconvincing
Critics note the benchmark is old, likely contaminated, and the harness broke its rules with tools and loops, with no comparison to harnessed frontier models.
Shared-file workflows already work
Several users already run similar plan/status/notes markdown setups with a single model, questioning whether the full manager-worker harness is needed.
Harness quality matters a lot
Commenters debate whether harnesses like Claude Code are holding models back, with most saying subagents work fine and the complainer likely has a config issue.

Consensus: Most agree the headline claim is weakly supported by the chosen benchmark, though mid-size open models like Qwen 27B are genuinely getting useful for local agentic work.

Recommended in the comments

Pi ↗ (tool): Praised for its compaction approach and running unattended, though with telemetry and dependency concerns.
llama.cpp (tool): Used with a detailed config to run Qwen 27B on a 24GB GPU.
OpenCode (tool): Called the best open harness one user has tried, with improving security.
Pullboard ↗ (tool): Its author built it for agent coherency on complex projects.
Cerebras ↗ (service): Reported 1500 tok/s for Qwen 27B, addressing speed concerns.

Summarised by Claude from the comment thread 1 day ago.

Points: 160
Comments: 56
Best rank: #1
Posted: 2026-10-25 by rzk
Category: Ai ml

multi-agent-orchestration qwen livecodebench llm-benchmarks code-generation open-weight-models python

All captures:

Full page

Captured 1 day ago. Click for the whole screenshot.

Looks similar

Screenshot of github.com: Show HN: Making a smart TV dumb and then smart

Show HN: Making a smart TV dumb and then smart (github.com)

69 points by animalhot 3 days ago | 38 comments

GitHub repo for r2d2-tv, a local home screen, private remote browser, and push-to-talk voice assistant that replaces vendor software on rooted LG webOS TVs.

webos smart-tv lg voice-assistant

Screenshot of github.com: Jukebox-OS: Deterministic time-travel replay and ACID tool sandbox for AI agents

Jukebox-OS: Deterministic time-travel replay and ACID tool sandbox for AI agents (github.com)

3 points by mariuscarlsson 21 hours ago | 0 comments

GitHub repository for JUKEBOX-OS, a Rust and Python runtime offering deterministic replay, a copy-on-write ACID tool sandbox, and time-travel debugging for multi-agent systems.

ai-agents time-travel-debugging sandboxing rust

Screenshot of github.com: Hedra is a 3D open world game I wrote from scratch while in highschool

Hedra is a 3D open world game I wrote from scratch while in highschool (github.com)

219 points by maxilevi 3 days ago | 41 comments

GitHub repository for Project Hedra, a low-poly procedurally generated 3D open-world RPG and custom C#/OpenGL engine written by its author starting in high school.

game-development game-engine csharp opengl

Screenshot of github.com: Show HN: Omnara – Open-source alternative to Claude Managed Agents

Show HN: Omnara – Open-source alternative to Claude Managed Agents (github.com)

85 points by kmansm27 4 days ago | 34 comments

GitHub repository for Omnara, an open-source, self-hostable managed agent platform for running durable AI agents with sandboxes, alternative to Claude Managed Agents.

ai-agents omnara durable-execution claude-managed-agents

Screenshot of github.com: PyTorch-native TPU back end for vLLM

PyTorch-native TPU back end for vLLM (github.com)

4 points by zniturah 6 days ago | 0 comments

GitHub repository for vllm-torchtpu, a PyTorch-native TPU backend for the vLLM LLM inference engine.

vllm tpu pytorch llm-inference

Screenshot of github.com: One Must Fall 2097 Remastered

One Must Fall 2097 Remastered (github.com)

38 points by hmm37 2 days ago | 21 comments

GitHub repository for a TypeScript/WebGL2 remaster of the 1994 robot fighting game One Must Fall 2097, with HD art, mods, LAN play and desktop/browser builds.

one-must-fall-2097 fighting-game remaster typescript