Liquid AI · 3 reports
| Engine | Avg t/s | Range | N |
|---|---|---|---|
| LM Studio | 129.0 | 129–129 | 1 |
| llama.cpp | 113.0 | 113–113 | 1 |
Unknown GPU · custom engine · 128,000 ctx
Reported by the source; GPU count, offloading and concurrent requests can change this figure. Check the full setup before comparing.
Running on OnePlus 13, pure CPU. Custom inference engine built from scratch, 450kb, supports other model architectures.
RTX 4050 6GB · LM Studio · 256,000 ctx
Reported by the source; GPU count, offloading and concurrent requests can change this figure. Check the full setup before comparing.
User benchmarked 20 small models on RTX 4050 6GB. Selected LFM2.5-1.2B-Instruct as cheap always-on model. Also tested Granite 4.1 3B, Gemma-4-agentic-e2b, Nemotron-3-Nano-4B, LFM2.5-8B-A1B. Full results table included.
AMD Strix Halo 128GB · llama.cpp · 128,000 ctx
Reported by the source; GPU count, offloading and concurrent requests can change this figure. Check the full setup before comparing.
Vendor benchmark on Ryzen AI Max+ 395. Also reports 30 tok/s on phone and 220 tok/s on M5 Max. Model is 2.69B params, 128K context, tool calling. Benchmarks: ToolSandbox 77.83, IFBench 59.17, BFCLv4 56.88, LiveCodeBench 59.41. Not recommended for agentic coding.