Results

Benchmarks

All numbers on this page were measured on one low-end reference PC. A faster machine gives higher speed. Pass rates depend on the model, not the machine.

CPU	AMD Ryzen 3 3200G, 4 cores / 4 threads (2019)
RAM	16 GB DDR4-2133, dual channel
GPU	none used (GT 710 is too old)
Disk	7200 rpm hard drive
Engine	llama.cpp CPU build, 4 threads, Q4_K_M models

Speed and quick tests

Speed from llama-bench (prompt 512 tokens, generate 64 tokens, 4 threads, CPU only). Quick tests: 8 single-function coding tasks, 10 vulnerability classes to identify, 5 small tool-building tasks.

Model	Prompt tok/s	Generate tok/s	Coding	Vuln ID	Tools
Qwen3.5-0.8B	142.20	33.32	3/8	9/10	1/5
Qwen2.5-Coder-1.5B	65.24	19.86	4/8	9/10	2/5
MiniCPM5-2B	44.88	15.24	8/8	10/10	4/5
LFM2.5-8B-A1B	50.35	13.20	8/8	not run	1/5
Llama-3.2-3B	33.29	11.03	6/8	10/10	2/5
Qwen3.5-4B	24.76	7.10	6/8	10/10	4/5
Gemma-4-E4B	20.71	6.79	8/8	10/10	3/5
WhiteRabbitNeo-V3-7B	15.89	5.17	5/8	10/10	3/5
Ministral-8B	14.69	5.00	4/8	10/10	4/5

Notes

Only completed runs are shown. A model without measured scores is left out.
"not run" means the test did not finish for that model, so no score is given.

Run it yourself

Run the same tests on your machine:

litecode tune                 # speed
litecode eval                 # all 30 challenges
litecode eval --level hard    # one level
litecode eval --kind security # one kind

Results are saved in <data_dir>/bench/.