Many developers were excited when Cerebras announced their hosted Qwen3 Coder service. They promised high speeds and a good alternative to Claude. But as time went on, it became clear that the reality didn’t match the hype. Users started to see that the promised performance and features weren’t quite delivering.

Promises vs. Reality: The Speed and Cost

At first, Cerebras claimed that Qwen3 Coder could handle 2000 tokens per second (TPS). That sounded impressive. For $50 or $200, you could get a plan that supposedly offered this speed. But when users tried to test it, they found out the speed was much lower. Many reports, including one from YouTube creator Adam Larson, showed that the actual TPS was often under 100, sometimes even less. Larson pointed out that he never hit the 2000 TPS mark, even in short tests. Instead, most of the time, the model was slow and unreliable.

This mismatch between the advertised and actual performance was frustrating. Users who rely on fast and consistent output for autonomous coding or other tasks found the service lacking. It turns out that the speed claims were overstated, and the infrastructure couldn’t deliver what was promised. This led to disappointment and questions about whether the service was worth the cost.

Technical Challenges and User Experience

Getting Qwen3 Coder to work smoothly wasn’t easy. One user reported that the command-line interface (CLI) didn’t even work at first. When they contacted Cerebras support, they blamed the user’s code, even though the same setup worked fine with other providers like Fireworks, Claude, and Alibaba’s hosted Qwen. The support team didn’t acknowledge the issues right away, and fixing the problems involved a lot of workaround coding. This was a stark contrast to Fireworks, which responded quickly and fixed problems almost overnight.

Another issue was the limited context window of just 131,000 tokens. For coding tasks, this is barely enough to handle complex projects. Users had to manage prompts carefully and keep their code small. Larger context models like Claude Code recently increased their window, but Qwen3 Coder remains constrained. This limitation affects how well the model can plan and generate autonomous code, often requiring users to step in and guide the process more than they’d like.

Cost, Limits, and Performance Comparison

The costs also didn’t line up perfectly. A Max account from Cerebras costs about four times more than a Pro account, but the token limits don’t scale proportionally. For example, a Max account allows 120 million tokens per day, while a Pro account offers 24 million. But the per-minute token limits are close, with Max allowing 400,000 tokens and Pro 300,000. This means that at peak times, users hit rate limits quickly and get errors that can interrupt their work.

Many users in Cerebras’s Discord chat were unhappy about the small context window and the rate limits. They also noticed that the performance of Qwen3 Coder on Cerebras seemed slightly worse than what they experienced with other providers or the official hosted versions from Alibaba. Larson’s review noted about an 8% drop in performance compared to the original models, which is noticeable when trying to do complex coding tasks or autonomous generation.

Despite these issues, Cerebras is still developing its system. They created their own Model Context Protocol to try to improve performance and reliability, but it’s clear that the service is still a work in progress. Many users are waiting to see if future updates can fix the speed, stability, and usability problems. For now, it seems that Cerebras’s Qwen3 Coder isn’t quite ready to replace more established options for AI coding or autonomous tasks.

Inspired by

• https://www.infoworld.com/article/4055909/down-and-out-with-cerebras-code.html

Sources

• gmpg.org
• unsplash.com
• cerebras.ai
• cerebras.ai
• github.com