The video reviews the new stealth AI coding model Sonic, highlighting its exceptional speed and decent performance on simple coding tasks but noting its limitations in conversational ability, complex reasoning, and content flexibility. The presenter speculates that Sonic is likely a Mistral coder variant optimized for speed, based on its large context window and behavioral traits, while inviting viewers to share their own theories.
The video introduces a new stealth AI coding model named Sonic, which was released recently and tested extensively in real production code. The standout feature of Sonic is its remarkable speed, reportedly generating a few hundred tokens per second, potentially up to 500, living up to its “Sonic” name. While the model performs adequately in coding tasks, it is not exceptional, and the presenter expresses curiosity about its long-term potential. Sonic is not expected to surpass current leading models like Claude but could be useful for specific tasks where speed is critical and the requirements are well-defined.
Several clues about Sonic’s identity are discussed, including its massive 262k context window, which is rare and shared by only a few models like Mistral, Quinn 3 coder, and Grocoder. The presenter leans towards it being related to Mistral coder rather than Grocoder or Anthropic’s Quinn, citing Sonic’s fast speed, limited conversational abilities, and strict content moderation as key indicators. Sonic is not very conversational and struggles with generating extended or creative responses, often refusing to produce inappropriate or mean content, which contrasts with the more flexible personality of Grocoder.
In terms of coding capabilities, Sonic shows mixed results. It can produce functional code for simpler projects like a pool game or a webOS desktop environment, with some success in physics simulations and UI design. However, it struggles with more complex tasks such as Unity 3D game development, often failing to produce working implementations despite multiple attempts. The model also shows limitations in conversational coding assistance, lacking the interactive, back-and-forth style typical of pair programming models.
The presenter highlights Sonic’s strengths in speed and some practical coding tasks but notes its weaknesses in reasoning, conversational interaction, and handling complex or nuanced programming challenges. The model’s strict content filtering and inability to engage in casual or off-topic dialogue further differentiate it from other AI coding assistants. Despite these limitations, Sonic’s rapid token generation and decent performance on straightforward coding tasks make it a potentially valuable tool for specific use cases.
Ultimately, the presenter speculates that Sonic is likely a Mistral coder model, possibly a smaller, expert-tuned variant optimized for speed rather than conversational depth. This theory is based on the context window size, behavior patterns, and performance characteristics observed during testing. While acknowledging a high chance of being wrong, the presenter invites viewers to share their thoughts and predictions, expressing excitement about the possibility of Sonic being a new Mistral release and its implications for the AI coding landscape.