Is Sonic AI a New Mistral Coder Model? An In-depth Review of Its Speed, Coding, and Conversational Abilities

GosuCoder

Introduction to Sonic AI [00:00:00]

A new stealth AI coding model, named "Sonic," was recently released and is notable for its speed.

• Initial Impressions: The model is "freaking fast," generating hundreds of tokens per second (estimated 200-500 TPS).
  • It is "all right" at coding but "not phenomenal."
  • It is not expected to overtake established models like Claude in its current version and configuration.
  • Its speed makes it useful for specific tasks where execution needs are clear.

Key Characteristics & Clues [00:00:55]

The presenter analyzes several characteristics of Sonic AI to deduce its potential origin.

• Context Window: It boasts a 262,000 token context window.
  • This is similar to Groq 4 (256k context) and Mistral Medium 3.1.

    

• Speed (Tokens per Second):
  • Demonstrates very fast token generation, estimated at over 200 TPS, possibly up to 500 TPS.
  • This is considered its biggest advantage.
• Conversational Ability: The model is "not conversational at all."
  • "Thinking Blocks" Issue: In platforms like OpenCode, its "thinking blocks" (internal reasoning steps) often contain only a single line, suggesting minimal internal processing or reasoning.
  • Storytelling Attempts: It provided very brief answers when asked for a "funny story" and required explicit instruction to generate a "long story" to produce more than two sentences.

    

• Language Capabilities:
  • Initially refused to tell a story in Chinese, stating it's "designed to assist with software engineering tasks."
  • However, it successfully generated code for a Tetris game and replied in Chinese when specifically prompted to do so.
  • It claims to be able to reply in multiple languages including English, Spanish, Chinese, Japanese, and Korean.

    

• Temperament (Inability to be "Mean"): The model consistently refused to generate inappropriate or negative content.
  • Refused requests for "dirty stories" or "lies."
  • Even when asked to "call me an idiot" or "insult my neighbor," it responded with kind, helpful, and polite replies, suggesting strong guardrails against harmful output.

    

Coding Performance Benchmarks [00:04:40]

The video showcases several coding tasks performed by Sonic AI, highlighting its strengths and weaknesses.

• Pool Game Simulation:
  • Temperature 1.0: Produced a non-functional game that stopped instantly.
  • Temperature 0.6: Generated the most functional version, although visually "ugly." It had working physics and balls could be pocketed, despite initial placement issues.
  • Temperature 0.75: The pool balls were placed incorrectly, and the cue ball could not be moved.
  • Optimal Temperature: The best performance was found between 0.6 and 0.8 temperature settings.
  • Low Temperature Issues: Below 0.5, it frequently resulted in tool call failures and significantly worse outputs.

    

• Modal Refactoring:
  • Successfully refactored an existing ugly modal in a codebase.
  • It performed in-line edits without reformatting the entire file, which is a positive sign for cleaner code integration.
  • The updated modal had a significantly better, more organized design with collapsed sections.
  • This task required "a lot of back and forth" iterations to resolve errors.

    

• Unity 3D Test:
  • Task: Create a jetpack function where holding space bar uses a power meter to jump higher.
  • Result: Constant failures and "dumb stuff," with numerous errors. Even after simplifying the prompt, the function did not work.
  • Conclusion: A "big failure" for complex 3D game development tasks.
• Interactive Sandbox Simulator:
  • Task: Create a physics sandbox.
  • Result: While it created basic shapes (polygons, rectangles, circles), the physics were flawed. Polygons and rectangles fell through the floor, and circles sometimes "spazzed out" when stuck together.
  • It included presets like a pendulum and Rube Goldberg machine, but these also did not function as expected.

    

• Website Creation ("Elite Code Architect"):
  • The model created a visually appealing website with a good design, glows, and subtle hover effects, especially for a non-designer.
  • Generated at 0.7 temperature.
  • Logic Failure: Key interactive elements like "View My Work" or "Get in Touch" buttons were not functional.
  • This was the best out of five versions, with others ranging from "god awful" to just "okay."

    

• WebOS Desktop Environment:
  • A "decent" but "not exceptional" environment was created.
  • Features included a basic notepad (text field in a large window) and a functional calculator that correctly performed multiplication.
  • Windows could be opened and overlapped.
  • This task also required multiple iterations and had initial console errors, indicating it was not "one-shotted."

    

• Arena Top-Down Shooter:
  • The game was not playable; the player character died instantly, and shooting mechanics were absent.
  • However, the controls felt good, suggesting potential if the core shooting functionality could be implemented.
  • Described as "Flash level" quality.

Prediction for Sonic AI's Origin: Mistral-Coder [00:10:59]

The presenter strongly predicts that Sonic AI is the long-anticipated Mistral-Coder model.

• Reasons for Prediction:
  1. Mistral's Absence: Mistral has not released a dedicated coder model, unlike other major players.
  2. Context Window Alignment: The 262k context window aligns with Mistral Medium 3.1.
  3. Model Temperament: The consistently "kind" and censored responses are characteristic of Mistral's general behavior, contrasting sharply with Groq's more flexible personality.
  4. "Thinking Blocks": The minimal, single-line "thinking blocks" suggest either low reasoning or a very efficient, coding-focused Mixture of Experts (MoE) model.
     • It is hypothesized to be an MoE model smaller than Qwen 3 Coder (Q3C 480b), specifically optimized for coding tasks.
     • The speed discrepancy with Mistral Medium's reported throughput could be explained if it's a smaller MoE or a paired release with a fast inference partner (e.g., Cerebras, Groq).
• Models Ruled Out:
  • Groq: Unlikely due to its known ability to be less censored and more conversational, unlike Sonic.
  • Anthropic and Qwen: No specific evidence points to these.
• Confidence Level: The presenter acknowledges a high chance of being wrong (95% chance), but the "feeling" strongly points to Mistral-Coder.