Two major model releases today: Opus 4.7 and Qwen 3.6.
I run an 18-task agentic coding benchmark that takes 3-5 hours. Real software engineering work - tool use, multi-file edits, complex reasoning. Qwen 3.6 is by far the best self-hosted model I’ve tested. It scores better than state-of-the-art 120B parameter models. I get 230 tok/s on my 5090 with 262k context and VRAM to spare.
It scores similarly to Gemini 2.5 Flash. Not “good for local” but actually competitive with cloud providers, with tool-calling support good enough to use it for real software engineering work.
This is the first self-hosted model that actually works.
the numbers
Qwen 3.6 on consumer hardware:
- 230 tokens/second on a 5090
- 262k context window
- Beats previous 120B models
- Matches Gemini 2.5 Flash performance
- Full tool-calling support
There are no rate limits, and nothing leaves your premises, so you never have to wonder whether your code is training someone else’s model. It’s just performance on hardware you own.
the other release
Opus 4.7: roughly 2-4% higher-scoring than 4.6 on my benchmark. Costs 4.5x as much per task.
The cache_create pricing is 4x higher. It builds more context per turn and uses a ton of thinking tokens. You’re paying for slightly better performance on problems that require extensive algorithmic reasoning - the hardest 10% of tasks where that extra reasoning depth matters. For everything else, you’re lighting money on fire.
the comparison
| qwen 3.6 (local) | opus 4.7 (cloud) | |
|---|---|---|
| performance vs previous | beats 120B models | 2-4% over 4.6 |
| token speed | 230 tok/s | rate limited |
| context window | 262k | varies by tier |
| cost | gpu amortization | 4.5x previous |
| data privacy | complete | none |
| tool calling | excellent | excellent |
The cost curve is going the wrong direction for cloud. Each marginal improvement costs far more than the last one, to the point where “better” now means dramatically more expensive for a small gain.
what this means
The two releases point in opposite directions.
One: AI becomes increasingly centralized and expensive. You pay more for less improvement, your data flows through someone else’s servers, you work within their rate limits.
The other: AI becomes distributed and sufficient. You run models good enough for real work on hardware from Microcenter. Performance improves with each release while staying on consumer hardware.
Qwen 3.6 is the first local model where I don’t miss the cloud for most engineering tasks. There’s still a capability gap - Opus 4.7 does handle those hardest 10% of problems better. That gap will probably always exist.
But when a consumer GPU matches Gemini 2.5 Flash, it’s worth asking why we’re paying 4.5x more for marginal gains. Running production-quality AI on the GPU you bought for gaming is now a real option rather than a compromise.