This model is better than ChatGPT and 10x cheaper

A new open-source “frontier” language model, DeepSeek V3, is being positioned as a major cost-and-capability shift: it reportedly cost about $5 million to train—far below the tens of millions to $100 million often associated with top-tier models—while delivering strong performance in everyday uses like English, coding, and math. The practical implication is that high-end model capability is moving out of the “only a few well-funded labs can afford it” category and toward a world where smaller startups can realistically build and iterate on competitive systems.

DeepSeek V3’s training approach is described as unusually selective rather than “grab everything from the internet.” Instead of a broad, low-curation scrape, the model was trained on a focused corpus of high-quality tokens, with explicit attention to performance across English, Chinese, math, and coding. After that, it was reinforced using human responses to improve accuracy—an ingredient that matters because it increases confidence during generation. That confidence, in turn, enables more efficient token prediction at inference time, letting the model generate ahead rather than relying strictly on one-token-at-a-time continuation.

The model’s efficiency is framed as a key technical differentiator. While DeepSeek V3 is described as a very large model (about 617 billion parameters), it does not use the full parameter space for every response. Instead, it draws on roughly 37 billion parameters—presented as a “sliver” of the total—for a given query. The claim is that this selective use of model capacity helps keep compute and latency manageable without sacrificing quality.

Another highlighted capability is multi-token prediction. Rather than predicting only the next token, DeepSeek V3 is said to predict two tokens ahead, leveraging the confidence built during training. The transcript also points to a training technique called “dual pipe,” described as a mechanism that allows learning and “regurgitating” (reusing learned patterns) to happen in a coordinated way through a specialized network setup.

Strategically, the emphasis is less on one-off benchmarks and more on what the cost and openness enable. DeepSeek V3 is open-sourced, meaning developers can inspect it, use it directly, and attempt their own improvements. That openness, combined with the much lower training cost, is portrayed as accelerating the spread of competitive “frontier” models.

The transcript contrasts this with a broader industry trend: the cutting edge is shifting toward inference-time compute—systems that run multiple parallel next-token prediction threads to find better continuations. In that framing, ChatGPT’s advantage is tied to inference-time compute, and open-source replication of that exact approach may take longer. Still, the immediate takeaway is that a $5 million model is outperforming or matching leading models (including ChatGPT-4 and Claude variants) on many common tasks, suggesting that “freeing” intelligence for business applications is becoming increasingly feasible—especially as replication gets cheaper and faster than the first breakthrough.