stable-code-instruct-3b-4bit
7 workload results across 3 hardware configurations.
Fastest local config
356.1 decode tok/s
on RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB via llama.cpp. see full run
For a standard 4-bit quantization using MLX, here are measured results for short-chat generation.
On MLX 4-bit, stable-code-instruct-3b is fastest on the M3 Ultra (60-core GPU) at 192.5 tok/s, 894% faster than the M3 Pro (18-core GPU) (19.37 tok/s).
| Card | Decode tok/s | TTFT (127-token prompt) | Run |
|---|---|---|---|
| M3 Ultra (60-core GPU) | 192.5 (single run) | 226 ms | r_y2_5y8oo97d |
| M3 Pro (18-core GPU) | 19.37 (single run) | 967 ms | r_pqjsvd-cub4 |
stable-code-instruct-3b, 4-bit, MLX, chat-short, batch 1. Only cards with their standard memory are listed; CPU-only hosts are excluded.
Local runs (7 workload results)
Runs from contributors' own machines via MLX, llama.cpp, vLLM, exllamav2, or ollama. Signed on the submitter's hardware.
RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | llama.cpp | - | 341.9tok/s | no data | 175ms | r_q9f15lz6831 |
| chat-long | llama.cpp | - | 289.1tok/s | no data | 232ms | r_q9f15lz6831 |
| concurrent-decode | llama.cpp | - | 356.1tok/s | no data | no data | r_q9f15lz6831 |
| agent-trace | llama.cpp | - | 292.8tok/s | 44,387.5tok/s | 42.7ms | r_q9f15lz6831 |
| chat-short | llama.cpp | - | 331.0tok/s | no data | 36.7ms | r_8l57cim1i10 |
M3 Ultra (60-core GPU) + 96GB unified
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | mlx@0.31.3 | - | 192.5tok/s | 560.7tok/s | 226ms | r_y2_5y8oo97d |
M3 Pro (18-core GPU) + 36GB unified
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | mlx@0.31.3 | - | 19.37tok/s | 131.3tok/s | 967ms | r_pqjsvd-cub4 |