Kilo Code x MiniMax M3 Live Webinar: Model Integration Demo

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

1. Buy a coding plan with Kilo credits

   Kilo's new coding plan marketplace lets you spend the same credits you use for inference through the Kilo gateway on a partner coding plan. The launch plan was MiniMax's token plan plus at about $20 per month, which the Kilo host describes as roughly 1.7 billion tokens of M3 per month (their figure, as stated in the webinar).

2. Route the repetitive, high-volume work to M3

   The MiniMax guest recommends M3 for high-volume long-horizon jobs such as code review and refactors, and keeping a heavyweight closed model for the hardest problems. For lower latency he suggests turning thinking off.

3. Use Kilo's code review tool as the harness

   Kilo's own benchmark, as described in the webinar, planted 17 issues in a buggy app: M3 caught 13 for about $0.07, while comparable runs on other models cost as much as $3 to $4. The guest credits both the model's low cost per token and the harness.

4. Diversify providers

   Both speakers argue for not locking into one provider as seat-based plans move to token billing (they cite GitHub Copilot's switch to usage-based billing and Anthropic's enterprise plans). Open weights let you run the model yourself if you have the hardware.

Gotchas

• Everything quantitative here is a vendor or partner claim from a co-marketing webinar; none of the benchmark numbers are independent.
• MiniMax describes M3 as a 1M-token context, natively multimodal model using block-level sparse attention (MSA), claiming about a twentieth of the per-token cost at long context and 9x faster prefill and 15x faster decode than M2. Treat these as unverified.
• The guest said open weights would be released on Hugging Face that Friday and that the model is larger than M2, so local hardware needs may be higher; he gave no VRAM figures.