Qwen3.8-27b-noslop-v2 (GGUF)
A mild "less slop" edit of Qwen3.8-27B, made with Heretic (Arbitrary-Rank Ablation) using the upstream config.noslop.toml recipe unchanged (trial 18 of 60). This is not a "no slop" model and it does not remove em dashes. It is published because, unlike our first release, it stays close to the original model in coding-agent runs.
Files (all built from the same edited weights with our own importance matrix; the multi-token-prediction head is retained in each):
| File | Size | Note |
|---|---|---|
Qwen3.8-27b-noslop-v2-Q2_K_XL.gguf |
9.83 GB | Same per-tensor layout as Unsloth's current UD-Q2_K_XL. The only file that was agent-tested (below). |
Qwen3.8-27b-noslop-v2-Q2_K_M.gguf |
10.9 GB | Stock llama.cpp Q2_K. Untested. |
Qwen3.8-27b-noslop-v2-Q3_K_M.gguf |
13.5 GB | Stock. Untested. |
Qwen3.8-27b-noslop-v2-Q3_K_XL.gguf |
13.1 GB | Unsloth UD-Q3_K_XL per-tensor layout. Untested. |
Qwen3.8-27b-noslop-v2-Q4_K_M.gguf |
16.8 GB | Stock. Untested. |
Qwen3.8-27b-noslop-v2-Q8_0.gguf |
29.0 GB | Stock. Untested. |
mmproj-F16.gguf |
0.93 GB | Vision projector, unchanged from Unsloth's. Image input was not tested. |
What was measured
All numbers are for this Q2_K_XL file against a base Q2_K_XL built the same way (same layout, same importance matrix).
| Check | Original | This model |
|---|---|---|
| Replies with fiction slop words (100 story prompts, greedy) | 86 / 100 | 65 / 100 |
| First-token KL divergence vs original | 0 | 0.0075 |
| Coding-agent runs passed (pi and Claude Code, 30 tasks x 2 rounds = 120 runs) | 110 / 120 | 106 / 120 |
| Runs stopped by a loop or no-progress detector | 2 / 120 | 6 / 120 |
| Average time per run | 34 s | 38 s |
| Average tool calls per run | 7.9 | 9.8 |
| Final agent message contains an em/en dash | 100 / 120 | 92 / 120 |
Read this honestly:
- The slop reduction is small (86 to 65 of 100, on fiction word lists only).
- It still loops a bit more than the original (6 vs 2 flagged runs of 120) and passes slightly fewer tasks. It does not meet a strict "never loops more than the original" rule.
- Dashes in agent final messages barely change (92 vs 100 of 120). Do not expect dash-free output.
- Agent test details: sandboxed tasks with hidden tests, 600 s wall limit, 100 tool-call cap, loop detectors. The test is small and the original itself fails or loops occasionally, so differences of a few runs are within noise.
- Only the Q2_K_XL quantisation was agent-tested. The other quantisations were built the same way but not evaluated; their loop behaviour is unknown.
If you need the strongest dash reduction in plain chat, our first release (RefinedNeuro/Qwen3.8-27b-noslop-gguf) does more, but it loops much more in coding agents (20 / 120 flagged runs in the same test). V2 trades effect for safety.
Usage
llama-server -m Qwen3.8-27b-noslop-v2-Q2_K_XL.gguf --jinja -c 32768
Works with the same chat template and sampling settings as Qwen3.8-27B. The multi-token-prediction head is retained.
Method
Heretic master with the unmodified upstream config.noslop.toml (KeywordRate over fiction slop words, first-token KL, ARA rank 50), 60 trials, TPE after 20 random starts, 4-bit search. Trial 18 was picked as the low-divergence candidate. The adapter was merged into the original bf16 weights and quantised. No tokens are banned and no output filtering is applied.
Related: other candidates from the same search were rejected after the agent test (for example trial 25, which scored well in the search but passed only 5 of 42 agent runs).