Qwen3.8 Flash Next vs Qwen3.8 27B in OpenCode, full context, on a 12 GB RTX 3060
Both models ran on one gaming PC (Intel i5-12400, 64 GB RAM, NVIDIA RTX 3060 12 GB), offline, as the model behind the OpenCode agent: Flash Next through Strata at 256K context, Qwen 27B through llama.cpp at 262K. Same prompt: a space shooter split into nine files. The agent could only read and edit files in its project folder. Builds that needed a fix say so, and the untouched original is here too.
The project prompt
A complete browser space shooter in nine files (index.html, css/style.css, js/config, input, entities, game, ui, audio, README): three enemy kinds, shield and triple-shot power-ups, a boss every five levels, best score saved, WebAudio sound effects, then re-read every file and fix anything that would stop the game.
The run recorded in the video: OpenCode on Strata, full 256K context. Works first time, untouched. Its own review pass fixed three bugs before it finished.
The first valid run of the same prompt, off camera. Also works first time, untouched.
Exactly as Qwen 27B wrote it at 262K context, thinking off: the title screen never starts. input.js expects onStart(fn) to be called, game.js overwrites it.
Qwen's build with two marked lines in js/game.js (search VITOMIND FIX): the two callbacks are now called instead of assigned. Then it plays well.
Speed on the RTX 3060 (our measurements)
One conversation at a time. Numbers are from this PC and these settings, not a universal benchmark.
| Model and settings | File · context | Tokens per second |
|---|---|---|
| Qwen3.8 Flash Next · Strata · 'Leave 20% for my work' (default) | IQ3_S · 256K | 6.5-8.3 |
| Qwen3.8 Flash Next · Strata · 'Full performance' | IQ3_S · 256K | 32-33 |
| Qwen3.8 Flash Next · inside OpenCode, while recording the screen | IQ3_S · 256K | ~19-26 |
| Qwen3.8 27B · llama.cpp · KV q8_0 in RAM · empty chat | GSQ-RCO IQ3_XXS · 262K | 5.3 |
| Qwen3.8 27B · same · 32K tokens in the chat | GSQ-RCO IQ3_XXS · 262K | 2.1 |
| Qwen3.8 27B · KV q8_0 on the GPU (context limited to 32K) | GSQ-RCO IQ3_XXS · 32K | ~20 |
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