VQV Signal

ROBOTS & HARDWARE WATCH TECHNICAL

Kintsugi 27B claims to fit Qwen3.8-27B on 12 GB GPUs

A post on X by @CaryPalmerr announced Kintsugi 27b, which the author says keeps the agentic scores of Qwen3.8-27B while supporting the full 262K-token context on 12 GB graphics cards. The author cites an AppWorld score of 294/336, compared with 293 for full BF16 precision. The post also says Kintsugi works in Claude Code, Codex, Pi and OpenCode and is released under Apache-2.0 or MIT. These figures are the author's own and have not been independently checked, and the post is mainly a video.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-21T23:45:00+00:00 Detected 2026-11-22T01:03:37+00:00
View original source

A post on X by @CaryPalmerr announced Kintsugi 27b, which the author says keeps the agentic scores of Qwen3.8-27B while supporting the full 262K-token context on 12 GB graphics cards. The author cites an AppWorld score of 294/336, compared with 293 for full BF16 precision. The post also says Kintsugi works in Claude Code, Codex, Pi and OpenCode and is released under Apache-2.0 or MIT. These figures are the author's own and have not been independently checked, and the post is mainly a video.

AI-assisted summary based on the listed source.

@CaryPalmerr announced Kintsugi 27b, claiming it keeps Qwen3.8-27B's agentic scores (AppWorld 294/336 versus 293 for full BF16) and the full 262K-token context on 12 GB cards, and works in Claude Code, Codex, Pi and OpenCode under Apache-2.0/MIT. These figures are the author's and weren't independently checked;...

If the claims hold up, owners of mid-range 12 GB GPUs could run a 27B model with a full long context window on their own hardware. Developers who use the named coding tools would be the most directly affected. Independent benchmark results are the thing to watch before relying on these numbers.

Hardware and robotics watchers may want to track whether this becomes a product, benchmark, or deployment signal.

Signal Strength 88% Technical label WATCH Public Interest 35 Category ROBOTS & HARDWARE Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 0 Practical Impact Score 26 Novelty Interest Score 94 Consequence Score 18 Curiosity Score 32 Shareability Score 50

VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).