Gemma 4 puts efficiency on the table: small models are starting to take business away

The Open-Source Efficiency War Forces Everyone to Make a Choice

Simon Willison posted an impromptu poll, asking developers to choose between Gemma 4 and Qwen 3.5. This isn’t just a reputation test—it exposes a fork in the road for open-source AI: small, practical models are challenging the old story of “the more parameters, the better.” After Gemma 4 was released on March 25, 2025, the conversation spread quickly. The topic shifted from “scale” to “whether it can be deployed.” For enterprises, this is very practical: when inference costs rise sharply, whether it can run reliably on affordable hardware starts to shape decision-making.

• At the data level: Gemma 4 has roughly 7B parameters, and it scores 82.5% on MMLU, directly undermining the assumption that “bigger is stronger”—especially when compared with larger Qwen 3.5 models, which need heavier GPU clusters.
• Ecosystem signal: Jeff Dean publicly acknowledged Gemma 4’s market feedback; developers have verified that it can run on consumer-grade hardware, and a consensus that “efficiency = competitiveness” is starting to form.
• Points of contention: Compared with Qwen’s long-context advantage, Gemma has also been questioned on its long-context performance. Also, while ZetaChain’s case of completing integration in one day is attention-grabbing, on-chain AI is still a niche scenario, so it can’t change the overall landscape.

My take: Efficiency is rewriting the logic of choice—whether you can complete deployment at low cost and with low barriers is becoming the top threshold for enterprise adoption.

• Developer preference for migration: Early users moving from closed subscriptions to self-hosted open-source weights value customization and cost reduction.
• Google’s expansion: Open-source “good-enough” small models force competitors to catch up on efficiency, or enterprise users will churn.
• Scale tailwind is shrinking: If players like Qwen can’t quickly make up for efficiency optimizations, the scale advantage will diminish at the margin in most real-world applications.

The Cost Ledger of “Scale vs. Efficiency”

Around Willison’s tweet, two interpretations emerged: one believes Gemma 4 is Google’s defensive move against its open-source push in Asia; the other argues it’s not really “frontier-level.” But what truly determines the direction of the industry isn’t the label—it’s the engineering signal that can be reused:

• ZetaChain’s report says it can achieve 81% KV-Cache compression in long-context scenarios, suggesting efficiency improvements may close capability gaps faster;
• In the supply chain layer, U.S. export controls on AI chips make “efficient, hardware-agnostic” models a hedging option;
• The metrics debate masks a direct consequence: lowering deployment barriers will accelerate enterprise-side POCs and limited rollouts, and there may be an explosion of AI-native applications before 2027.

Key point: Efficiency creates a systemic premium. In the short term, it benefits small teams that can iterate and deliver quickly—and it’s also forcing a reassessment of the “giant models first” path.

Conclusion: Models like Gemma 4—“lightweight and usable”—are forcing the real cost into the open. Efficiency-first players will complete the conversion from PoC to production faster.

• Significance：High
• Categories：Model Release, Industry Trend, Open Source

My view: Investors and builders betting on the “efficiency narrative” are still early and currently in a favorable position. The real beneficiaries are delivery-oriented Builders and enterprise-side solution teams. If you’re a strategy-driven fund that only bets on “parameter scale,” this narrative isn’t friendly for short-term trading; but for funds and industry M&A making mid-to-long-term allocations, it’s worth resetting positions.