A $4,699 DGX Spark can easily run Qwen 3.6 35B, and its performance crushes GPT-o1, the SOTA from two years ago
373 karma · joined March 20, 2026
A $4,699 DGX Spark can easily run Qwen 3.6 35B, and its performance crushes GPT-o1, the SOTA from two years ago
China, a "authoritarian state" country, "the antonym of freedom", with a software industry that is especially capitalist, has produced all the competitive open-weight models.
It really is IRONIC.
Disclosure: I am Chinese, and I understand this strategy comes from being behind, using open source as an asymmetric way to compete and make up for missing compute by sharing the burden, etc. But still, very ironically.
https://www.youtube.com/watch?v=RWwGFDynOHo
For these basic virtual car experiments, BeamNG.drive is a pretty good physics simulator. You can open its built-in tools and run braking tests directly.
https://www.bilibili.com/video/BV1fME16uEW7
If the time-to-first-token latency also greatly improved, this could be very useful for end-to-end in controls, like autonomous driving for example.
Two years ago, SOTA was gpt-o1, and it was much more expensive than Fable. Now, for $4,699, you can easily run a much smarter Qwen3.6-35B locally with DGX Spark.
Think about where we are. This is an era where a new SOTA arrives every two months. It took LLMs only about 18 months to go from chain-of-thought reasoning to disproving the unit-distance conjecture. chatGPT itself is only three and a half years old.
DeepSeek V4, released two months ago, is almost as cheap as the electricity costed, has the ability to being absolutely a top-tier model in 2025 standards.
One possibility is that software starts to look more like traditional manufacturing.
The machine is the company’s core asset. The engineer only needs to know how to operate the machine well. Once that happens, the barrier gets much lower, need much less people, and the job naturally become much less valuable. Some parts will still need to be done by hand, of course. But only a very small part. It is like old factories. They used to need lots of fitters, at all levels of skill. Now you only need a few of the elite ones.
AI is the CNC machine of the software industry.
The more pessimistic future is that, maybe five years from now, the best programmers will look at AI the same way the best Go or chess players look at AI today: Like KeJie said, "I don't even know what I am trying so hard against." We now have a new SOTA every two months. It just took 18 months for LLMs from reasoning models to disproving the unit distance conjecture. ChatGPT itself has not even existed for as long as a college student spends in university.
In any case, we have already passed the point where this can be rolled back.
Maybe ten years from now I will be leaving a comment saying that "programmer" used to be a job too :-/
Programming is the low-hanging fruit for AI. Open source and knowledge sharing have given it huge amount of public, high-quality training data at a level other industries can hardly imagine. And almost everything in programming can be tested and verified inside the computer quickly in a closed loop. No robot arm is needed.
The main weakness of current LLMs is still that they are static: They do not really change themselves through use. Harness tools are just elaborate ornamentation on top of prompts. LLMs are frozen at the moment training stops. Once we get models that can change their own weights through self-feedback, then maybe AGI really is on the horizon.
Thinking optimistically: I may be lucky enough to see it in my lifetime. Maybe by then, people will be able to live more like human beings, instead of organizing their whole lives around work :-)
There is still a lot of optimization left to squeeze out of LLMs.
I also don’t think the extra $20 will hurt GTA’s sales much. What I’m more worried about is whether other 3A games will use Rockstar’s price increase as an excuse to raise their own prices . And I’m even more worried that GTA may come out with a pile of DEI slop, which would really feel like another Atari crash.