Descriptions:
Fahd Mirza puts Inkling-Small through its paces in a hands-on evaluation — testing the newly released 276-billion-parameter mixture-of-experts model from Thinking Machines across four distinct tasks. Despite the “Small” branding, the model activates 12 billion parameters per token, carries a 1-million-token context window, natively handles text, images, and audio, and ships under an Apache 2.0 license.
The practical tests include: debugging a FastAPI app with Redis Pub/Sub (the model correctly identified mismatched channel names and produced a working one-line fix on the first attempt), generating a rotisserie chicken simulation in a single self-contained HTML file (functional but visually average compared to top competitors), translating an audio-delivered sentence into nearly 80 languages (solid for most, some literal or off), and parsing a form image into structured JSON (clean section-by-section extraction with no missed fields).
Mirza’s overall verdict is nuanced: vision and audio capabilities are genuinely strong, code generation is adequate but not exceptional, and multilingual performance is competent rather than excellent. He also highlights a notable finding from Thinking Machines’ own benchmarks — at every thinking-effort setting, Inkling-Small’s performance curve sits above the larger Inkling model on coding, HULA exam tasks, and instruction following. For teams currently using the bigger Inkling model, the efficiency case for switching is compelling.
📺 Source: Fahd Mirza · Published July 30, 2026
🏷️ Format: Review
