Liquid LFM2.5 24B vs Qwen3.8 27B on a 12 GB RTX 3060
Both models ran on one gaming PC (Intel i5-12400, 64 GB RAM, NVIDIA RTX 3060 12 GB) with llama.cpp, no internet, no API. Same prompt: a night-street e-bike game in one HTML file. One run each. Cards that needed a fix say so, and the untouched original is there too.
The game prompt
Night-time city street, three lanes, potholes and cones, battery packs, turbo, HUD, start and game-over screens. One HTML file, canvas only.
Liquid's build. It called a function it never wrote, so it showed a black screen: one marked line fixes it. Basic shapes, short runs.
Exactly as Liquid wrote it, unfixed: it stops on a missing function (updateBackground).
Qwen's build, untouched: parallax city, street lights, cones, potholes, batteries, charge bar and turbo. Written with thinking switched off.
Speed on the RTX 3060 (our measurements)
llama.cpp b11797, CUDA 13.4, one conversation at a time. Numbers are from this PC and these settings, not a universal benchmark.
| Model and settings | File | Tokens per second |
|---|---|---|
| Qwen3.8 27B, default launch (17.9 GB file spills into RAM) | Unsloth UD-Q4_K_XL | 2.8 |
| Qwen3.8 27B, all layers on the GPU | ISTA-DASLab GSQ-RCO IQ3_XXS | 18 to 21 |
| Qwen3.8 27B + MTP speculative decoding, thinking off | GSQ-RCO IQ3_XXS-mtp | 20 to 28 |
| Liquid LFM2.5 24B, default launch (14.5 GB) | LiquidAI Q4_K_M | 57 |
| Liquid LFM2.5 24B, all layers on the GPU | Unsloth UD-Q3_K_XL | 103 to 127 |
| Liquid LFM2.5 24B + DSpark drafter (one conversation at a time) | UD-Q3_K_XL + DSpark F16 | 29 to 75 (vs 101 to 127 without it) |
The exact prompts are for Ruocco Academy members. The Academy is not open yet: join the waiting list and you get one email when it opens. The demos on this page are free to use for everyone.
Join the waiting list