Summary
David Ondrej interviews Ditro, co-founder and CTO of Fireworks AI, for a wide-ranging conversation on the state of open-source AI models. The central claim: models like Kimi K3, DeepSeek V4 Flash, and Gemini 5.2 have collectively delivered to open weights the kind of agentic capability jump — longer-horizon tasks, sub-agent coordination, goal-directed loops — that closed frontier labs achieved in late 2024. Ditro groups these releases together as a single step-change comparable to when Opus 4 first made agentic coding viable.
The conversation then turns to the economics of model customization. Ditro argues that full pre-training from scratch is becoming the province of fewer and fewer organizations due to cost and scale requirements, while post-training and fine-tuning represent a high-leverage on-ramp for companies wanting to specialize open weights for their domain. Fireworks AI has built its platform around co-designing inference and training infrastructure to serve this middle tier of customers. He uses Bloomberg GPT as a cautionary tale — a domain-specific pre-train that was quickly surpassed by general-purpose models — to argue that specialization through post-training is a more durable equilibrium.
The video also includes a sponsored segment for Appodex’s Frontier Agent, an open-weights 35B parameter agentic framework that runs locally on a MacBook, demonstrated generating a data-backed chart about open-source model quantization timelines using a multi-agent verification pipeline.
📺 Source: David Ondrej · Published September 09, 2026
🏷️ Format: Interview
