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-egirl-LoRA
hemlang/Hemlock-Qwen3.6-27B
Input Modalities
Output Modalities
Supported Functionality
| axis | cases | TIES | Vorarbeiter | Wichtel |
|---|---|---|---|---|
| delegate | 10 | 0/10 | 7/10 | 10/10 |
| no_tool (restraint) | 10 | 9/10 | 9/10 | 10/10 |
| trivial (one-line edits) | 5 | 4/5 | 5/5 | 5/5 |
| inspect | 5 | 5/5 | 5/5 | 5/5 |
| exec | 4 | 4/4 | 4/4 | 4/4 |
| git | 4 | 4/4 | 4/4 | 4/4 |
| memory | 3 | 3/3 | 3/3 | 3/3 |
| web | 3 | 3/3 | 3/3 | 3/3 |
| persona | 3 | 3/3 | 3/3 | 3/3 |
| total | 47 | 35/47 | 41/47 | 47/47 |
47/47 valid tool calls, zero hallucinated tool names. Greedy decoding, so these are deterministic.
Every generated program is executed by the real interpreter and its stdout compared against expected output. Not string matching.
| level | TIES | Schierling/TIES | Wichtel | |
|---|---|---|---|---|
| L1 | syntax | 1/9 | 8/9 | 7/9 |
| L2 | stdlib | 0/5 | 1/5 | 2/5 |
| L3 | algorithms | 0/7 | 5/7 | 4/7 |
| L4 |
No code agent understands Hemlock. Vorarbeiter's delegation policy hands engineering work to a coding agent, which is correct for every language that agent knows and wrong for this one — if the operator does not write Hemlock itself, nothing does. So this model has to do both: delegate ordinary engineering, and write Hemlock directly.
The interesting result is that the two capabilities did not fight. The risk was concrete: the Hemlock adapter is 5,562 rows of direct code generation with zero tool calls, exactly the data that could teach an operator to write code instead of delegating. Instead delegation went from 7/10 to 10/10.
An earlier attempt to get both from a single ORPO run failed in a specific, measurable way: folding 400 Hemlock preference pairs into the delegation data dropped restraint from 9/10 to 4/10, because all 400 rows call a tool and nothing in the training set teaches answering directly. Splitting the objectives — SFT for the language, preference training for the behaviour — fixed it and taught the language far better besides (400 preference pairs moved hembench approximately nothing; 5,562 supervised rows moved it 55 points).
markdown
Qwen/Qwen3.6-27B└─ huihui-ai/Huihui-Qwen3.6-27B-abliterated└─ nbeerbower/Qwen3.6-27B-TIES 5-way TIES merge└─ + Bubba-3ep (GreatFirewall ORPO — states facts it otherwise sanitises)└─ + delegation LoRA (hands engineering work to a code agent) = Vorarbeiter└─ + Schierling (Hemlock SFT, 5562 rows) = Hemlock-Qwen3.6-27B└─ + egirl LoRA (delegation + Hemlock tool use, ORPO) ← Wichtel
Four adapters deep, all four capabilities intact.
peft merge_and_unload() silently drops the 15 mtp.* tensors and current llama.cpp then
refuses to load the model. They are grafted back from the pre-merge base and verified
bit-identical. 1,199 tensors total (850 language, 333 vision, 15 MTP, 1 lm_head).
bash
llama-server -m Wichtel-Q8_0.gguf -ngl 99 -c 8192 --jinja -fa on -np 1 \--spec-type draft-mtp --spec-draft-n-max 4
MTP speculation is not bit-identical in practice — on this bench one decision flipped between speculative and non-speculative decoding at temperature 0, apparently from numerical differences between batched verification and sequential decoding flipping a near-tie.
code_agent, glob_files, git_status…).
A different tool set needs different training data.nbeerbower/Hemlock-Qwen3.6-27B — this model without the egirl adapternbeerbower/Schierling-Qwen3.6-27B-LoRA — the Hemlock adapternbeerbower/Qwen3.6-27B-egirl-LoRA — the tool-use adapternbeerbower/Vorarbeiter-Qwen3.6-27B — the operator base| systems |
| 0/7 |
| 5/7 |
| 6/7 |
| L5 | translation | 0/5 | 3/5 | 3/5 |
| L6 | debugging | 3/5 | 4/5 | 4/5 |
| overall | weighted | 10.1% | 65.0% | 67.6% |
| code that fails to run | 30 | 8 | 7 |