GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
teilomillet
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
teilomillet
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
f7bdf119fb573577756da28978e39d28bcfdb5b000f8982c3ebff6faaaafc1cfeb069009f424ac5f851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0aa8504dc05c3151818ff9988847c220be730242a34e0d406b42e1f8d11344008aTraining initially failed at step 38 with microbatch size 2. It restarted at microbatch size 1 with gradient checkpointing from adapter weights only; optimizer, scaler, and RNG state were not restored. This weakens exact-continuation claims and is intentionally disclosed.
Same 8 tasks, four rollouts each, seed 123, 64 turns, 512 action tokens, temperature 0.8, top-p 0.95, no oracle:
| Candidate | Passed |
|---|---|
| Raw base | 0/32 |
| Previous adapter | 10/32 |
| Step 628 | 16/32 |
| Step 1256 | 21/32 |
| Step 1884 | 19/32 |
| Final | 24/32 |
The final checkpoint was the unique gate leader and was selected before Spider evaluation.
The single scoreable run on the predeclared 16-task dev split scored 0/16 with the official scorer. All 16 tasks terminated on repeated-action detection and none submitted a result. Therefore this artifact is a negative benchmark result despite its factory-gate improvement. Do not infer leaderboard competitiveness from the factory score.
The companion private evaluation repository contains the frozen selection record, transcripts, official score, diagnostics, taxonomy, preflight, and operational incident logs. No Spider heldout task was used for training or checkpoint selection.
Under the 32-episode factory gate, the selected checkpoint submitted 27/32 episodes with 24 positive rewards, 4 repeated-action cutoffs, 0 token-cap hits, and 0 turn exhaustions. On the 16-task Spider dev split, all 16 episodes ended in repeated-action cutoffs with no submission.
This adapter is an agent policy for the quaero openenv_v1 dbt/DuckDB tool contract (list_files, read_file, query, build_spec, write_file, dbt, validate, submit). It emits JSON tool actions and expects environment observations in return; it is not a general chat or one-shot text-to-SQL model. Generate with enable_thinking=False.
The environment implementing this contract is open source at github.com/teilomillet/quaero-dbt — including UserWorkspaceRuntime for driving this adapter against your own dbt project, and the verifier that mirrors the official Spider 2.0-DBT scoring semantics.
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelBASE_REV = "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", revision=BASE_REV)tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-4B", revision=BASE_REV)model = PeftModel.from_pretrained(base, "teilomillet/quaero-qwen35-4b-sft-factory-longhorizon-20260711")
The full story — why this adapter climbs the synthetic factory gate and still scores zero on real Spider 2.0-DBT instances, and what that measures — is written up at teilo.xyz/posts/quaero-spider2-negative-result.