GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ginaecho
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
ginaecho
Model Tree
Qwen/Qwen2.5-Coder-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
V7 achieved 10/10 joint validity/faithfulness passes on ten development role-name variants of one synthetic workflow family, using 2,206 native construction tokens. A fresh-process confirmation reproduced those aggregate quality and token results. Mean semantic F1 was 93.59, not 100.
However, on four full project intents:
| Case | nuScr | Semantic F1 | Generation calls | Construction tokens spent |
|---|---|---|---|---|
| Finance | Invalid | Not assessable | 6 | 15,301 |
| AgenticPay settlement | Valid | 17.73 | 1 | 391 |
| React 18 migration | Invalid | Not assessable | 6 | 8,920 |
| SDLC concurrent release gate | Invalid | Not assessable | 6 | 12,896 |
Full-workflow joint passes: 0/4. Invalid syntax prevents a complete semantic inventory in three cases; unavailable F1 is not zero. AgenticPay passes static validation but sends goods and payment in the wrong directions. None of these outputs hit the generation-length cap.
Do not use a generated draft as an accepted protocol without independent static validation and intent assessment. nuScr validity does not establish faithfulness or enforce value refinements. Low token expenditure is not evidence of a cheap, correct replacement for a skills-based workflow.
The exact prompt retains the pilot's payload-free restriction. Rich workflow intents can require payloads and refinements beyond that training scope. Changing the prompt, generation budget, engine, precision, or runtime creates a different evaluation configuration and does not inherit the reported performance.
Use Linux, Python 3.13, and a CUDA GPU supporting BF16. The measured runtime used an NVIDIA H100 NVL with PyTorch 2.6.0/CUDA 12.4. A smaller GPU may work, but its behavior and memory requirements are not verified here. The base alone is about 15 GB. Do not run this on a machine without sufficient storage and GPU memory.
Install the PyTorch wheel matching that environment, then the remaining packages:
bash
python -m pip install torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124python -m pip install -r requirements.txtDownload only the small inference script from this repository, or clone the repository on your GPU machine. Then supply an intent text file:
bash
python inference.py --intent-file example_intent.txt --revision v7The script anonymously downloads only this release's adapter/configuration,
manifest, and system prompt. It obtains the base from the pinned upstream
revision. --base-local-files-only requires an already cached base. It refuses
CPU fallback, checks artifact hashes and the chat template, never truncates the
input, and returns JSON with the draft, exact native token counts, runtime
identity, and the resolved adapter revision.
For archival reproduction, replace v7 with the commit SHA recorded by your
run. v7 identifies this frozen release and will not be moved to a later model.
The script uses trust_remote_code=False.
The output is explicitly not validated. The published release_check.json
records an anonymous public-adapter reload and a pinned nuScr check on the small
development example. That smoke check is not another held-out benchmark.
| Field | Value |
|---|---|
| Base | Qwen/Qwen2.5-Coder-7B-Instruct |
| Base revision | c03e6d358207e414f1eca0bb1891e29f1db0e242 |
| Adapter weights SHA-256 | 8f13b0b28f65c57b327021b19a4f2861a7082e36f61446d1c791fcdb2c33226e |
| LoRA | Rank 32, alpha 64, dropout 0.05 |
| Target modules | q/k/v/o projections and gate/up/down projections |
| Selected lineage | Original checkpoint 15 plus a new two-optimizer-step continuation |
| V7 learning rate |
Most examples were template-authored structural tasks. Four additional successful-path examples were paraphrased from primary documentation for S3 multipart upload, TCP handshake, MQTT QoS2, and OAuth authorization-code exchange. Targets were admitted only after compiler and semantic qualification.
Training does not backpropagate through nuScr or the semantic reviewer. This is supervised specialization, not reinforcement learning. Repeated development feedback guided checkpoint selection. No broad generalization claim is supported.
The following numbers concern the ten synthetic development variants only. Construction stops at the first nuScr pass, with at most five repairs after the initial draft. Scoring costs are excluded. Mean F1 includes every scored valid draft, not only joint passes.
| Revision | nuScr valid | Joint passes | Mean F1 (n) | Calls | Construction tokens |
|---|---|---|---|---|---|
| V0, unadapted | 0/10 | 0/10 | N/A (0) | 60 | 19,620 |
| V1, fresh checkpoint 15 | 6/10 | 6/10 | 92.86 (6) | 30 | 9,357 |
| V2 | 10/10 |
The original in-process checkpoint-15 evaluation scored 7/10; fresh loading consistently scored 6/10. This unresolved discrepancy is preserved rather than replacing the fresh-loading reference with the higher score. V7 confirmation accepted a user-approved blank-line-only difference in one draft; the original strict byte-identical audit failed and remains part of the project evidence.
The adapter is released under Apache-2.0. The pinned Qwen base has its own
Apache-2.0 license and is not redistributed here. Original STJP prompt/code
materials retain the MIT attribution in NOTICE.
This release contains only the adapter, configuration, public model metadata, small inference example, aggregate results, and licensing information. It does not include credentials, private operational logs, Azure resource identifiers, or third-party source-document copies. The training data and nuScr Docker image are not included; their availability must not be inferred from this model release.
The formal image used for the recorded checks was
sha256:cee12ae9f33a91dc4e8b0bdae5d34813d563a6d0ae63af89ddb9b109d92cfb1c.
This digest is an identity, not a public registry download URL.
| 0.00005, cosine schedule, warmup ratio 0.1 |
| Batch / accumulation | 2 / 8 |
| Training pool | 104 qualified unique examples; 158 sampled entries |
| Objective | Completion-only supervised next-token loss, prompt labels masked |
| Training sequence bound | 2,048 tokens, no truncation |
| Inference | Greedy, BF16, SDPA, seed 20260922 |
| Inference bounds | 1,024 new tokens; 8,192 input-plus-output tokens |
| 0/10 |
| 75.62 (10) |
| 10 |
| 2,223 |
| V3, qualification only | N/A | N/A | N/A | N/A | N/A |
| V4 | 10/10 | 0/10 | 77.16 (10) | 10 | 2,190 |
| V5 | 10/10 | 0/10 | 76.58 (10) | 10 | 2,194 |
| V6 | 10/10 | 0/10 | 76.67 (10) | 10 | 2,194 |
| V7 | 10/10 | 10/10 | 93.59 (10) | 10 | 2,206 |