GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
loktar
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
loktar
Model Tree
Qwen/Qwen3.6-27B
Input Modalities
Output Modalities
Supported Functionality
vLLM (works on bf16, FP8, or AWQ base — all tested):
bash
vllm serve Qwen/Qwen3.6-27B-FP8 \--enable-lora --max-lora-rank 32 \--lora-modules fresh-paint=loktar00/Qwen3.6-27B-fresh-paint
Then request "model": "fresh-paint". System prompt used for all evals: "You generate a single, self-contained HTML document that fully answers the user's request." Sampling: temperature 0.7, top_p 0.95.
llama.cpp: stock convert_lora_to_gguf.py crashes on this hybrid-attention architecture (V-head reorder). Either use a patched converter that dispatches input-dim reorders to lora_A / output-dim to lora_B, or build merged GGUFs via llama-export-lora from a base GGUF + this adapter. Serving with --lora <converted>.gguf on a Q4_K_M base is well-tested.
Merging note: merge under a healthy transformers — a known save bug in 5.14.1 inflates merged saves of this architecture ~4x.
Headless-Chromium runtime audit (uncaught exceptions + blank-render detection), 94 prompts, served FP8+LoRA. fresh-paint = SFT epoch-2 + DPO (this release); e1 = SFT-only sibling:
| slice | fresh-paint | SFT-only (e1) | base |
|---|---|---|---|
| styled incl. held-out styles (60) | 60/60 | 54/60 | 54/60 |
| general, no style named (16) | 14/16 | 14/16 | 15/16 |
| games (10) | 10/10 | 9/10 | 10/10 |
| rare trained styles (8) | 8/8 | 8/8 |
Blind human preference (same 94 prompts, 3-way, position-randomized panels): fresh-paint 36 · e1 26 · base 20 (12 skips). On styled prompts the tuned pair took 85% of votes.
The honest paragraph. In a later blind against its own SFT parent (e2) at Q4, fresh-paint won the styled category (8-6) but the parent won overall (22-16), driven by games and general prompts — and a human playability pass found more of fresh-paint's game pages broken-in-play than the parent's, a defect class automated audits largely miss. Net: the DPO stage bought style fidelity and killed a fence artifact (below) at some cost to game robustness relative to its immediate parent. If your use case is games-first, benchmark both; the SFT parent adapter may be released separately.
```html fences on 39% of outputs (a trained-in teacher habit); the DPO in this release eliminated it — 0/94. If you parse outputs from any model in this family, prefer document-anchored extraction (find <!doctype → </html>) over fence regexes.max_tokens caps. /no_think soft switch works as in base.Apache-2.0, inheriting the base license. Training data is teacher-distilled synthetic output from commercial frontier models plus preference pairs from the model's own drafts; no benchmark content. Trained and evaluated by loktar.
| 8/8 |
| total /94 | 92 | 85 | 87 |