Qwen3.8 finetunes

The post-train ecosystem around Qwen model family’s 27B dense — the model the doombox runs as its default (llama-swap (OSS) Qwen38-27B-dflash2-think). Finetunes are chosen by what they change: behavior (uncensored), efficiency (thinking budget), or packaging (quants). The reasoning-efficiency branch matters most for the local setup: thinking tokens are the dominant latency and VRAM cost of a 27B at 245k context.

Finetunes

• OrcaRouter Qwen3.8-27B-Uncensored (OrcaRouter) — abliterated; behavior change, not performance. On the box as Orca38 (bartowski Q4_K_M). See Uncensored models for the school.
• GestaltLabs Ornstein3.6-27B (Gestalt Labs) — SFT → NSC-ACE (GRPO + latent steering) → SABER; the surgical-edit school. On the box as Ornstein36 (3.6-generation base).
• UkisAI Swift 1.5 Qwen3.8-27B — the reasoning-efficiency flagship (model card captured): 58.5% fewer thinking tokens, +0.35% aggregate score, 9.18× speed-up; the demo game build fell from 104.6 to 11.39 minutes. Training method is the interesting part: identify tokens linked to pathological overthinking and penalize those — without attacking reasoning length directly — then recover accuracy with RL + OPD (on-policy distillation); data re-sampled from their multi-turn agent SFT set into proper RL environments. Gains concentrate exactly where the local profile works: LiveCodeBench 76.76→81.71, Terminal-Bench 2.1 69.21→72.13 (at −16% tokens); math takes a small hit (AIME 98.67→96.00). Swaps into llama-swap as a plain GGUF — drop-in Qwen38 replacement candidate.

Reduced-thinking finetunes

The sub-school that attacks overthinking directly — distinct from quantization (same tokens, cheaper each) and from speculation (DFlash2: draft-and-verify, exact same output). Reduced-thinking finetunes change how many thinking tokens get emitted, trading a measured sliver of accuracy for large latency/context savings:

• Swift 1.5 (UkisAI) — the current reference: token-level penalty on overthinking patterns, RL+OPD recovery; median-token reductions 47-58% on general reasoning, near-zero on agentic loops (Terminal-Bench median −0.1% — agentic chains resist compression, an honest limit).
• ThinkingCap (BottleCap AI, ThinkingCap-Qwen3.8-27B) — imatrix + per-tensor precision route: 37% fewer reasoning tokens at −0.8 points; “no file is shown to be lossless” honesty in reporting. Same goal, quantization-flavored approach.
• Swift 1.5 Flash-Next edition — the same recipe applied to the 125B MoE (Qwen model family’s Flash-Next section), bundled by Strata (OSS) as its default pick — short thinking scales to the big model too.
• Base models’ own budget controls (--thinking-budget-style flags) are the zero-training alternative; finetunes win when the model overthinks pathologically (looping, hedging) rather than just long.

Composability question flagged for the box: does Swift 1.5 stack with the DFlash2 drafter? Both target the same profile slot (Qwen38), both report lossless/near-lossless claims, but nobody has published the combination — worth a hands-on A/B: Qwen38-27B-dflash2-think vs Swift 1.5 plain vs both.

Related: Qwen model family · ThinkingCap-Qwen3.8-27B · DFlash2 · llama-swap (OSS) · Local LLM inference · Unsloth