Mozilla State of Open Source AI V1.0 Report

Open-Weight Models Reach Capability Parity

Open-weight AI models have effectively closed the capability gap with closed-source frontier models for most production workloads. As of March 2026, the average capability gap on Chatbot Arena stood at 3.3%, with open models reaching parity in coding, instruction-following, and general knowledge. The remaining gap is concentrated in complex reasoning, long-context retrieval, and agentic tasks.

この収束はインテリジェンスコストの崩壊とともに起きています。GPT-4 クラスの推論コストは 36 ヶ月で 27 倍低下し、1M トークンあたり $20 から $0.40 にまで下がりました。その結果、OpenRouter などのプラットフォームではトークン量でオープンウェイトモデルが支配的となり、上位 5 つのモデルはすべてオープンウェイトです。2026 年中頃までに、中国製オープンモデルは週あたり約 18T トークンを処理し、米国製モデルの約 5.5T と比較して大きく上回っています。

The Operational Gap: Adoption vs. Production

While open models lead in adoption, they lag significantly in production deployment due to a lack of operational tooling. According to a Mozilla/SlashData 2026 survey:

• Adoption: 79% of developers adding AI functionality use open models, compared to 71% for closed models.
• Production Rate: Only 51% of open-model teams reach production, whereas 63% of closed-model teams do.

このギャップは企業規模に関係なく続いています。従業員 1,001 人以上のエンタープライズ規模チームでは、クローズドモデルの生産率が 73% に対し、オープンモデルは 57% にとどまります。主な阻害要因は能力ではなく運用面で、既存システムへの統合、継続的な保守、デプロイの複雑さが挙げられます。

The Shift to the Agentic Harness

As model weights commoditize toward zero cost, value is shifting upward to the "agentic harness"—the orchestration loop, memory, sandboxes, and permission models that turn a model into a functional agent.

The Integration Moat

Frontier labs are increasingly welding their harnesses to their weights to create a performance moat. Data from Terminal-Bench 2.1 shows that lab-owned harnesses consistently outperform independent scaffolds. This vertical integration creates a data flywheel where usage through the lab's harness directly informs the next model iteration, increasing vendor lock-in.

The "Write Surface" Problem

A critical unsolved gap in the open stack is the "write surface"—the portable permission model that defines which irreversible actions (e.g., spending money, modifying records) an agent can perform unattended. Current protocols like MCP (Model Context Protocol) and A2A focus on authentication (who the agent is) rather than authorization (what the agent may do).

Open AI as a Sovereignty Strategy

Open-weight AI is increasingly viewed as a matter of national sovereignty rather than a simple vendor choice. Over 70 national AI strategies are now active, driven by the need for "exit rights" to avoid proprietary API traps.

• Risk of Shutdown: The report cites a June 2026 incident where a government export order forced Anthropic to cut access to Claude Fable 5 for all foreign nationals instantly, highlighting the fragility of relying on closed endpoints.
• China's Strategic Lead: China has intentionally used open weights as a macro hedge against semiconductor export controls. By releasing public weights (e.g., Alibaba's Qwen), China offloads global inference onto end-users' local hardware. As of March 2026, Qwen downloads significantly outnumbered the next eight organizations combined.

Commercial Viability of the Open Stack

Open-weight AI is a multi-hundred-billion-dollar commercial market. Proven revenue models include hosted inference, enterprise platforms, on-prem licensing, and harness tooling. Key players include:

• DeepSeek: Raised $7.4B at a $50B+ valuation with ~$220M ARR.
• Mistral AI: Scaled to ~$400M ARR with a valuation around $14B.
• Databricks: Reached a $5.4B revenue run-rate.

Community Perspectives and Critiques

Discussion surrounding the report highlights a tension between the "open-weight" reality and the "open-source" ideal.

"I am sad that there doesn't seem to be any community whatsoever around truly open models that are released with source data and training methodology... We've allowed the term 'open' to be diluted to a shocking extent."

Other critics pointed out that the report's prose appeared to be LLM-generated and that the reliance on OpenRouter data may be biased, as users of that platform are already predisposed to seek alternatives to frontier closed models. However, some observers noted the explosive growth of open models, with one user reporting a 5× increase in open-model token processing on OpenRouter over just four months.

---

要約: Mozilla の 2026 年レポートは、オープンウェイト AI モデルがコーディングや一般知識においてクローズドモデルにほぼ同等の性能を達成し、競争の焦点が「エージェンシーハーネス」や運用ツールへとシフトしていることを示しています。

タイトル: Mozilla State of Open Source AI V1.0 Report