Qwen 27B Runs Full Web App on Dual 3090s — Open-Source Catches Closed Models

A real-world test on Reddit this week made us stop and take a closer look: Alibaba's community Qwen 27B model completed an 80s-style retro web game called RetroCraft in one shot within a few hours, running on two consumer-grade RTX 3090 GPUs. The output included complete HTML, JS, and CSS code packaged as a runnable application, with an average generation speed of ~71 tokens/sec (tokens being the smallest text units a model processes). The signal is clear: Chinese open-source large models have hit visible, usable performance on consumer hardware — local AI coding is leaving the geek circle behind.

What this is

The author used a fully local setup: two RTX 3090s (used market price around 2,000–3,000 RMB each, the 2020 consumer flagship) + the llama-swap model scheduler + the OpenCode open-source coding assistant.

Three key details:

First, the scale: 27B (27 billion parameters), Q8 quantized (compressed from high precision to 8-bit to slim it down). This is the borderline "barely fits" point for consumer GPUs.

Second, the author kept the model's "thinking process" enabled (thinking tokens — the internal scratchpad used during reasoning), letting the chain of thought run through the entire task. Results were noticeably better than the version with this disabled — a relatively new open-source community practice over the past six months.

Third, speed swings widely: from 10 to 464 tokens/sec within a single turn. The 71 token/sec average is workable, but it won't always be smooth.

Industry view

Community feedback has been broadly positive. 71 tokens/sec sits in the "good enough" range — humans read code at roughly 5–10 lines/sec, so AI output is no longer the bottleneck.

But the caveats deserve airtime:

One-shot success doesn't equal stable reproducibility. Coding agents (AIs that can autonomously plan and execute multi-step tasks) still need prompt tweaking and retries on complex jobs.

The hardware bar isn't low. Two 3090s plus a host machine runs 10,000–20,000 RMB total, and you need to understand CUDA configuration, quantization parameters, and inference tuning — a dealbreaker threshold for non-technical users.

"Qwen 3.8" isn't an official Alibaba version number — it's a community modded build. For a regular user wanting to reproduce this, the path is unclear and documentation and compatibility are shaky.

MTP (Multi-Token Prediction — generating multiple text units at once to speed things up) is a newer technique; its cross-task stability still needs more sample validation.

Impact on regular people

For enterprise IT: highly regulated sectors — finance, healthcare, government — now have a "Chinese open-source + local deployment" double safeguard, no longer needing to ship code to overseas APIs. The total procurement cost for small and mid-sized IT teams is lower than many assume.

For working professionals: AI coding assistants may shift from tens of dollars per month in subscriptions (Cursor, Copilot) to a one-time hardware outlay. But reaching the plug-and-play moment for non-technical users is still 1–2 years of productization away.

For the consumer market: demand for consumer GPUs will rise, and prices on new releases like the RTX 5090 won't drop in the short term.