GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
vitorallo
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
vitorallo
Model Tree
Qwen/Qwen2.5-Coder-32B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Out of scope: standalone chat, non-security code tasks, dependency-CVE (SCA) analysis, and runtime/config issues that are not visible in source.
Injection (SQL, command, code/eval), insecure deserialization (RCE), path traversal,
SSRF, XXE, XSS, open redirect, broken access control, IDOR / broken object-level
authorization (BOLA), broken authentication (incl. weak password-reset tokens), CSRF,
sensitive-data exposure, and C/C++ memory-safety (buffer overflow, UAF, integer overflow).
Held-out test set (1,887 examples, served FP8 via vLLM):
| Metric | Value |
|---|---|
| Precision | 41.0% |
| Recall | 40.9% |
| F1 | 40.9% |
| False-positive rate (clean code) | 0.7% |
| JSON-schema compliance | 98.8% |
The very low false-positive rate (0.7%) is the headline characteristic — the model is calibrated to stay quiet on clean code rather than over-flag, which is what makes it usable in an automated pipeline. Recall (~41%) is per-function single-pass; scanner-level recall is higher because functions are reviewed with call-graph context and candidate rules.
Per-language recall (test set): C# 65%, Rust 53%, Ruby 52%, JS 45%, Python 44%, Java 43%, TypeScript 42%, Go 40%, C++ 39%, C 32%.
Qualitative check with the Foil scanner against DVNA (OWASP Top-10 reference app): the model identified the documented data-flow and access-control vulns — SQL/command/code injection, XXE (RCE), insecure deserialization (RCE), broken access control, IDOR/BOLA (with ownership-fix reasoning), the A2 MD5-of-login reset-token logic flaw, and unvalidated redirect.
Known limitations. Reflected XSS whose sink lives in a separate view template
(.ejs/.hbs unescaped output) can be missed when only the handler is in context — supply
template context for full coverage. The model is thorough and may surface advisory findings
(rate-limiting, headers) beyond a strict vuln set; triage accordingly. Dependency-CVE (A9),
runtime misconfiguration (A6), and logging (A10) are out of scope for source-only review.
This is a compressed-tensors FP8 checkpoint (FP8_DYNAMIC) — vLLM auto-detects the
quantization from the checkpoint, so no --quantization flag is needed:
bash
vllm serve vitorallo/securereview-coder-32B-FP8 \ --served-model-name securereview \ --max-model-len 32768The model emits a JSON object:
text
{"findings": [{"severity", "category", "line", "code", "description", "recommendation", "confidence", "cwe_id"}]}Quantization recipe: FP8_DYNAMIC (per-channel static weight scales + per-token dynamic
activation scales), lm_head kept in higher precision.
Built on Qwen2.5-Coder. Please cite the base model:
text
@article{hui2024qwen2coder, title={Qwen2.5-Coder Technical Report}, author={Qwen Team}, year={2024}}vllm serve vitorallo/securereview-coder-32B-FP8 \ --served-model-name securereview \ --max-model-len 32768