Open-Source AI’s Leap Forward: Linas Beliūnas Highlights DeepSeek’s Impact

In a recent LinkedIn post, Linas Beliūnas explores the seismic shift occurring in the artificial intelligence landscape, specifically highlighting the advancements made by DeepSeek and their implications for open-source AI development. Beliūnas posits that open-source AI may have not only caught up to but potentially surpassed the capabilities of proprietary, frontier labs.

Beliūnas opens by expressing astonishment at DeepSeek’s latest releases, V3.2 and V3.2-Speciale, suggesting that their performance metrics challenge the established hierarchy in AI development. He notes the surprising performance figures, stating:

“DeepSeek just launched V3.2 and V3.2-Speciale, and the numbers don’t look real: → Beats GPT-5 High → Rivals Gemini 3 Pro → Gold medals in IMO, CMO, IOI → 100% open-source (MIT license) → ~5× cheaper than GPT-5 (24× for output tokens)”

The author of the post emphasizes that an open-source model achieving such results is a departure from previous expectations, where open-source was anticipated to be significantly behind closed-door developments. Beliūnas points out that DeepSeek’s breakthrough wasn’t achieved through sheer scale, but through fundamental architectural improvements.

Rewiring the AI Engine

Beliūnas details the innovative approach DeepSeek has taken, moving beyond brute-force parameter expansion. He explains that the company has focused on optimizing the core engine of the AI. According to Beliūnas, key advancements include:

• Smarter gradient flow
• Deeper Reinforcement Learning (RL)
• A sparse-attention system designed to reduce costs and increase stability
• A training pipeline typically associated with private, cutting-edge labs

This sophisticated engineering, as described by Beliūnas, signifies a leapfrog rather than a mere catch-up in the AI race.

The Power of Raw Reasoning

A particularly striking aspect for Beliūnas is the performance of DeepSeek’s V3.2-Speciale model, which achieves remarkable results without relying on external tools or browsing capabilities. Beliūnas highlights this as a critical differentiator, stating:

“V3.2-Speciale doesn’t use tools. No browsing. No external calls. Just raw reasoning. And yet it still posts: – 96% AIME – 99.2% HMMT – Gold-level IMO + IOI performance – Top-tier coding benchmarks beating GPT-5 High”

This ability to perform complex reasoning and achieve top-tier benchmark scores without external assistance, according to Beliūnas, fundamentally redefines what is considered a ‘frontier’ AI capability. He argues that this level of performance from an open-source model equipped only with raw reasoning is a significant development.

Democratizing Advanced AI

Beliūnas concludes by outlining the broad implications of DeepSeek’s advancements. He contends that this development puts considerable pressure on closed, proprietary AI labs across several fronts: pricing, performance, transparency, access, and the pace of innovation. In Linas Beliūnas’s view, this opens up unprecedented opportunities:

“This pressures closed labs on pricing, performance, trust, access, and pace. → Startups suddenly get elite intelligence without elite budgets. → Researchers get transparency while developers get freedom. → Entire countries get capabilities once locked behind API gates.”

The core takeaway for Beliūnas is the potential shift of AI leadership from massive corporations to open-source communities. He suggests that the most powerful AI might soon emerge from the collaborative efforts of developers working rapidly, rather than solely from trillion-dollar companies. Beliūnas summarizes this paradigm shift by stating, “The frontier isn’t gated anymore. It’s finally forkable.” This suggests a future where advanced AI capabilities are more accessible and adaptable, driven by open collaboration.