Tess-4.5-27B for Splash

A Splash package of Tess-4.5-27B, a Qwen3.8-27B fine-tune, for LM Studio's Splash runtime on Apple silicon.

Credit

• The model: Tess-4.5-27B by migtissera. This package is only a re-packaging of their fine-tune into Splash format — all credit for the model itself goes to them. None of this would exist without their work.
• Splash engine, and the official Qwen3.8-27B Splash package this conversion reuses the DFlash 2 draft and vision tower from: Inco AI.
• Conversion tooling: publicExcess/splash-converter.

Converted from the BF16 checkpoint (revision 91b231133470bd147aac8036ed94d35e9bf1dc13) with the community splash-converter. The DFlash 2 draft and the vision tower are reused from Inco AI's official Qwen3.8-27B Splash package — the draft is the standard Qwen3.8-27B one, not retrained for this fine-tune. Package format: splash-packed-q4, schema 3, 78 artifacts, ~17.4 GB.

Measured speed

Mac Studio, Apple M5 Max, 128 GB. Identical ~60-token prompt, temperature 0, 1,000 max tokens, one model resident at a time, decode tok/s excluding prefill:

Use with LM Studio

Requires an M3 or newer Mac, macOS 26.4 or later, 36 GB memory minimum (48 GB+ recommended) and the Splash runtime (Settings → Runtime). Download this repo into your LM Studio models folder; it is cataloged as a Splash model and loads on the Splash engine.

Use with the standalone server

Current standalone Splash loads MLX checkpoints directly and no longer loads packages. Serve the MLX build instead:

brew install incoai/tap/splash
splash serve --model juno/Tess-4.5-27B-MLX-4bit

Notes

• The package carries no generation_config.json (the format has nowhere to put it); set sampling in your client. The /v1 API defaults to greedy decoding.
• Converted and measured by Juno for Duckets, 2026-10-19.