Arete-27B-Uncensored β€” Splash package

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

Credits

  • SargeDev β€” creator of Arete-27B-Uncensored: the abliteration of Arete-27B (aggressive SVD refusal-direction removal) that this package is converted from. The model weights here are his work.
  • Nous Research β€” creators of Arete-27B, the value-alignment SFT of Qwen3.8-27B that SargeDev abliterated.
  • Inco AI β€” the Splash engine and the official Qwen3.8-27B package, whose DFlash2 draft and vision tower are reused here byte-for-byte.
  • publicExcess β€” the open-source converter that built the target weights.

Conversion

  • Source: SargeDev/Arete-27B-Uncensored BF16 checkpoint, revision 47a9768a60861e5ba8aff363a1c205aaabb1a64d.
  • The source repo is a text-only export (Qwen3_5ForCausalLM layout). Before conversion, its language-model tensors were renamed back to the family's multimodal layout (model.language_model.*) β€” a pure rename, no values changed β€” and fitted with the base Arete config. Its 15 MTP (multi-token prediction) head tensors are not carried: the Splash package format has no MTP component, and speculative decoding is provided by the DFlash2 draft instead.
  • Target converted from the BF16 source 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, 17,382,657,779 bytes (~17.38 GB).

Usage

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

License

Apache-2.0, following the source model.

Verified in LM Studio

Smoke-tested in LM Studio on a Mac Studio (M5 Max, 128 GB) on 2026-11-08, loaded as arete-27b-uncensored-splash:

  • Reasoning β€” PASS: 391 / 13 answered with brief working (13 Γ— 30 = 390, remainder 1, β‰ˆ 30.077).
  • Vision β€” PASS: a solid-red 64Γ—64 PNG was correctly identified with the one-word answer "Red".
  • Speed: 77.8 tok/s measured via the local API (800 completion tokens in 10.28 s, default load settings). For reference, the original Arete-27B-Splash package benched 104.5 tok/s under the same settings.

The model was unloaded after testing.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Duckets/Arete-27B-Uncensored-Splash

Base model

Qwen/Qwen3.8-27B
Finetuned
(13)
this model