GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
nbeerbower
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
nbeerbower
Model Tree
nbeerbower/Qwen3.6-27B-delegation-LoRA
nbeerbower/Qwen3.6-27B-TIES
nbeerbower/Bubba-Qwen3.6-27B-LoRA-3ep
Input Modalities
Output Modalities
Supported Functionality
| TIES (base) | Vorarbeiter | |
|---|---|---|
| tool + args correct | 7/10 | 9/10 |
| delegated coding work | 0/3 | 3/3 |
| valid tool calls | 10/10 | 10/10 |
| ccp_truth (corroborate) | 18.00/18 | 18.00/18 |
| ccp_truth_neutral (volunteer) | 2.40/4 | 3.20/4 |
| compliance | 2.00/2 | 1.80/2 |
| safety_control | 2.00/2 | 2.00/2 |
| capability | 3.00/3 | 3.00/3 |
| censorship TOTAL | 27.40/29 | 28.20/29 |
| prose distance | 0.98 ± 0.07 | 1.05 ± 0.07 (not distinguishable) |
Censorship eval: 29 held-out items, 5 samples each at temperature 0.7, marker-scored. Behaviour
in this family is unstable run-to-run — the same prompt under greedy decoding can give a whitewash,
a refusal, or the truth — so figures are rates, not single passes. Tool-calling: ten operator
requests in egirl's <tool_call> dialect with its real system prompt.
This is the interesting part. Applied alone, the delegation adapter fixed tool use but knocked the censorship axis it never trained on:
| ccp_truth | neutral | TOTAL | tool+args | delegate | |
|---|---|---|---|---|---|
| TIES | 18.00 | 2.40 | 27.40 | 7/10 | 0/3 |
| + delegation only | 17.20 | 2.40 | 26.60 | 8/10 | 3/3 |
| + Bubba only | 18.00 | 3.60 |
Adding Bubba back recovered everything delegation displaced and pushed tool accuracy to 9/10 —
higher than either adapter alone. The over-eager web_research call that delegation-alone
introduced on an opinion question is gone too.
Contrast with a same-axis stack: censorship-direction orthogonalisation (28.00 alone) combined with Bubba scored below either component in all three orderings tried. Same-axis interventions subtract; cross-axis interventions add. That is the rule these numbers support, and it is why this model was built by stacking rather than retraining.
Merging the adapters reproduces the runtime-stacked result within sampling noise (28.20 vs 28.60 — one item at n=5).
peft merge_and_unload() silently drops the 15 mtp.* tensors — the multi-token-prediction head —
because they are not in the module tree transformers tracks. Current llama.cpp then refuses to load
the model (missing tensor 'blk.64.attn_norm.weight').
They are grafted back here from the pre-merge base. Since a language-model LoRA never touches that head, the copy is exact. The model therefore also supports MTP speculative decoding:
bash
llama-server -m Vorarbeiter-Q8_0.gguf -ngl 99 -c 8192 --jinja -fa on -np 1 \--spec-type draft-mtp --spec-draft-n-max 2
If you merge adapters into a Qwen3.6 model yourself, check mtp.* survived — this architecture has
1,199 tensors total (850 language, 333 vision, 15 MTP, 1 lm_head).
markdown
Qwen/Qwen3.6-27B└─ huihui-ai/Huihui-Qwen3.6-27B-abliterated└─ nbeerbower/Qwen3.6-27B-TIES (5-way TIES merge)└─ + Bubba-Qwen3.6-27B-LoRA-3ep merged (GreatFirewall-DPO, ORPO, 3 epochs)└─ + Qwen3.6-27B-delegation-LoRA merged (egirl-delegation-dpo, ORPO, 2 epochs)└─ + MTP head grafted back ← this model
Vision tower (333 tensors) carried through unmodified.
compliance dipped 2.00 → 1.80 — one circumvention question, one sample. Noise-level, but
reported rather than rounded away.code_agent, glob_files, git_status…). A
different tool set needs different training data.nbeerbower/Qwen3.6-27B-TIES — the basenbeerbower/Bubba-Qwen3.6-27B-LoRA-3epnbeerbower/Qwen3.6-27B-delegation-LoRAschneewolflabs/egirl-delegation-dpo| 28.60 |
| — |
| — |
| + both (runtime) | 18.00 | 3.60 | 28.60 | 9/10 | 3/3 |