Qwen 3.8 27B: Opus-level intelligence that runs locally on a 4090

Man, I love Chinese AI labs. Qwen 3.8 - a 27 billion parameter model - is now on the market, and in several benchmarks it has the intelligence of Opus 4.6, which was a 500B+ parameter model.

The benchmark shock

Typed out as a table, it is wild seeing a 27B dense model trade blows with the frontier:

Bench	Qwen 3.8 27B	Opus 4.6 Max
SWE-bench Pro	61.7	53.4
OSWorld	84.3	72.7
QwenSWEBench	79.0	63.8
AndroidWorld	81.9	62.0
LiveCodeBench	90.3	88.8

Same SWE benchmarks as Opus. Mind-blowing.

Capabilities

• 1M token context (262K native -> 1M via YaRN).
• Native vision-language - handles screenshots, video, text.
• Apache 2.0 - open license, run anywhere.
• Code Arena rank #9 overall - the top open Apache-2.0 model in the top 10.

Local hardware

Here is where it gets fun: with 4-bit quantization, Qwen 3.8 27B runs on an RTX 4090. Roughly 13-17 GB VRAM, ~41 tokens/sec. On hardware you already own.

The new paradigm

Compare the two worlds:

• Cloud frontier: $20-200/mo, rate limits, your data leaves the box.
• Local Qwen 3.8: free, ~37 cents per million tokens, single GPU, fully yours.

The frontier-vs-home intelligence gap is closing. I will take that look. Thank you, Qwen.