GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
freeqaz
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
freeqaz
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
This adapter comes out of a campaign teaching fleets of AI coding agents to reconstruct shipped game binaries byte-for-byte — matching decompilation: rewrite C++ until the original vintage compiler, at the original flags, reproduces the retail machine code exactly, with every proposal judged by the real compiler and objdiff, never by a model. Start with the flagship write-up and the methodology post; the full project context lives on the r32 card, which is the release to use for actual decompilation work.
Because the pair is a clean matched experiment and the result is a null:
across k=5, k=20 at T=0.3/0.6/1.0, and six repeat draws per condition, the
r64 arm measured within the repeat-draw noise floor of the r32 arm at every
operating point (e.g. clean solves 17-18 vs 17-18 at k=20 T0.6; +1 to +2 at
other points, all inside the measured gray band). Doubling adapter capacity
at fixed data bought nothing detectable at 9B. If you are studying
LoRA-capacity effects, this pair gives you both arms of the experiment with
the corpus difference provably zero — the training-job-written
PUBLISHED.json in each repo pins the same corpus sha256.
Use the r32 for actual decompilation work (same measured capability, half the adapter bytes). Model facts, prompt format, usage, integrity and training-data provenance are identical to the r32 card apart from r/α and the adapter size (660 MiB vs 330 MiB).
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelbase = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", dtype="bfloat16", device_map="auto")tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B")model = PeftModel.from_pretrained(base, "freeqaz/decomp-synth-lifter-v13-chain-qwen3.5-9b-r64-lora")
RAW dec prompt format inside a compile-and-score loop, exactly as
described on
the r32 card.
| adapter | base | format | role |
|---|---|---|---|
| v10 | Qwen3.6-27B | RAW | single-turn matcher, reference 27B |
| gtp0c | Qwen3.8-27B | RAW | v10 recipe verbatim on the newer base (matched pair) |
| v13-chain r32 | Qwen3.5-9B |
PUBLISHED.json (shipped) carries per-file sha256 written by the training
job at publish time. Two internal metadata files it pins
(train_summary.json, artifact-manifest.json) are retained in the
project's artifact store but not shipped here — they carry training-box
paths. adapter_config.json was rewritten post-training to the HF base id;
original bytes at provenance/adapter_config.orig.json.
Adapter weights: Apache-2.0. Same corpus, same provenance as the r32: public CC0-1.0-dedicated decompilation projects. See the v10 card for the full discussion.
| RAW |
| headline single-turn 9B |
| v13-chain r64 (this repo) | Qwen3.5-9B | RAW | rank-capacity null twin of the r32 |
| v14-evalformat | Qwen3.5-9B | chat | multi-turn feedback-driven, strongest agentic 9B |
| v16-v5chat | Qwen3.5-9B | chat | broadest corpus, research artifact with stated negatives |
| v17-full | Qwen3.5-9B | chat+tools | first tool-trajectory generation, unevaluated |
| v17-matched | Qwen3.5-9B | chat+tools | byte-exact-only trajectory filter of v17-full (contrast arm), unevaluated |
| v17-full r64 | Qwen3.5-9B | chat+tools | rank-capacity twin of v17-full r32 (r64/α128), unevaluated |
| v17-mix | Qwen3.5-9B | chat+tools | v17 trajectories + single-turn supply mix (broadest v17 corpus), unevaluated |