Duckets/Arete-27B-Splash

Arete-27B — Splash package

A splash-packed-q4 package of Nous Research's Arete-27B for the Splash engine (LM Studio's Splash runtime, or the standalone splash serve).

Credits

• Nous Research — creators of Arete-27B, the value-alignment SFT of Qwen3.8-27B this package is converted from.
• Inco AI — the Splash engine and the official Qwen3.8-27B package, whose DFlash2 draft and vision tower are reused here byte-for-byte.
• publicExcess — the open-source converter that built the target weights.

Conversion

• Source: NousResearch/Arete-27B BF16 checkpoint, revision 108f0afbdcfaa1ceeda4a635004304dfff6d4da2.
• Target converted from the BF16 original with the splash-converter (4-bit packed). The draft (DFlash2) and vision components are byte-copies of the family's official package, incoai/Qwen3.8-27B-Splash, as the draft is family-specific and the vision tower is config-identical across the fine-tunes.
• Package verified by the converter: 78 artifacts, 73 weight files, ~17.37 GB.

Verified in LM Studio

Loaded and smoke-tested in LM Studio (Splash runtime) on an M5 Max (128 GB):

• Reasoning check: 17 × 23 = 391, √169 = 13
• Vision check: solid-color image identified correctly
• Speed: 104.5 tok/s decode (single stream, ~60-token prompt, temp 0, default load settings). Loading with a very large context (e.g. 262K) benches lower — ~68 tok/s in a follow-up test with the same harness.

Usage

• LM Studio: download into your LM Studio models folder; it is cataloged as a Splash model (arete-27b-splash) and loads with the Splash runtime.
• Standalone: splash serve loads raw checkpoints directly — point it at NousResearch/Arete-27B; this package format is for the LM Studio runtime.

License

Apache-2.0, following the source model.