Needs review (flagged 2026-10-13): This 3B-model result (19.4 tok/s) is slower than 7B and 8B models measured on the same M3 Pro, which is unexpected. It is kept public while it is reviewed.
stable-code-instruct-3b-4bit on M3 Pro (18-core GPU) + 36GB unified
MeasuredOn 2026-04-28, stable-code-instruct-3b-4bit ran at 19.4 tok/s (chat-short) on an M3 Pro (18-core GPU) with mlx. Run r_pqjsvd-cub4.
M3 Pro (18-core GPU) + 36GB unified
suite suite-v1
cli 0.0.1-dev
signedS8711zXnps…
Embed badgesubmitted Apr 28, 2026Workload results
| Workload | Backend | Model | decode tok/s | prefill tok/s | TTFT | p50 | p95 |
|---|---|---|---|---|---|---|---|
| chat-short | mlx@0.31.3 | mlx-community/stable-code-instruct-3b-4bit | 19.37tok/s | 131.3tok/s | 967ms | 17.5ms | 45.1ms |
Reproduce on your machine
Start with this model and workload selection. This command does not capture the original artifact, backend version or server settings:
Match quantization, context capacity, caching and runtime settings for a useful comparison. Duration varies by workload and hardware; the published client may differ from this run's version. How it's measured.
Share this run
Copy the short link, a Reddit-ready summary, or a README badge. Every number is the measured result on this page, and each badge links back to it.
https://llm-speed.com/r/r_pqjsvd-cub4Related benchmarks
- More stable-code-instruct-3b-4bit benchmarks — every backend and rig that has run this model.
- More M3 Pro (18-core GPU) LLM benchmarks — every model measured on this hardware.
Provenance
- Run ID
- r_pqjsvd-cub4
- Fingerprint hash
- a52e5dd258afe436
- Public key
- S8711zXnpsbOS9F8EZCne0DE3jWiyeYAqEDECBzTVWk=
- Received
- 2026-04-28 14:00:02