AI code generators need large models that can manage a broader context window, able to handle around 100,000 lines of code. Mixture of expert (MOE) models designed for agentic and reasoning AI are also sizable. However, these massive models tend to be quite costly, with prices ranging from $10 to $15 per million output tokens on current GPUs. This presents an opening for innovative AI architectures to challenge the dominance of GPUs.

Cerebras Systems Launches Big AI with Qwen3-235B

Cerebras Systems (a client of Cambrian-AI Research) has introduced support for the substantial Qwen3-235B model, which supports a context length of 131K (approximately 200–300 pages of text), four times what was previously possible. At the RAISE Summit in Paris, Cerebras highlighted Alibaba's Qwen3-235B, which leverages a highly efficient mixture-of-experts architecture to achieve remarkable computational efficiency. The real breakthrough, however, is that Cerebras can operate this model at just $0.60 per million input tokens and per million output tokens—less than one-tenth the price of similar closed-source models. Although many view the Cerebras wafer-scale engine as expensive, this data challenges that belief.

A question I often receive is: if Cerebras is so fast, why doesn't it have more customers? One reason is that it previously lacked support for large context windows and bigger models. Developers looking to create code don’t want to break problems into smaller fragments to fit, say, a 32KB context. Now, this sales barrier has disappeared.

“We’re seeing significant demand from developers for cutting-edge models with extended context, particularly for code generation,” said Andrew Feldman, CEO and Founder of Cerebras Systems. "Qwen3-235B on Cerebras is our first model that competes directly with leading-edge models like Claude 4 and DeepSeek R1. And with full 131K context, developers can now utilize Cerebras for production-level coding applications and get responses back in under a second instead of waiting minutes on GPUs.”

Cerebras has increased its supported context length from 32K to 131K tokens—matching the maximum supported by Qwen3-235B. This enhancement significantly affects the model’s ability to process large codebases and complex documentation. While a 32K context suffices for basic code generation tasks, a 131K context allows the model to simultaneously handle dozens of files and tens of thousands of lines of code, enabling development for production-grade applications.

Qwen3-235B performs exceptionally well in tasks demanding deep logical reasoning, advanced mathematics, and code generation, thanks to its capability to switch between "thinking mode" (for high-complexity tasks) and "non-thinking mode" (for efficient, general-purpose dialogue). The 131K context length empowers the model to ingest and analyze large codebases (tens of thousands of lines), supporting tasks such as code refactoring, documentation, and bug detection.

Cerebras also revealed further growth in its ecosystem, gaining support from Amazon AWS, DataRobot, Docker, Cline, and Notion. The inclusion of AWS is especially significant;

Where is this heading?

Big AI has continuously been reduced and optimized, achieving significant improvements in performance, reductions in model sizes, and cost decreases. This trend will likely continue but will be counterbalanced by enhancements in capabilities, accuracy, intelligence, and completely new features across different modalities. Therefore, if you're satisfied with last year's AI, you're in good shape since it continues to become cheaper.

But if you seek the latest features and functions, you'll need the largest models and the longest input context lengths.

It's the Yin and Yang of AI.

The above is the detailed content of Who Needs Big AI Models?. For more information, please follow other related articles on the PHP Chinese website!