Mozilla: The State of Open Source AI (V1.0)

Date: July 2026 | Status: Recurring Report

---

✉️ A Message from Raffi Krikorian, CTO

The real-world impact of open-weight AI is already unfolding in the corners of the globe where proprietary "black box" models cannot reach:

• New Zealand: A Māori broadcaster is utilizing speech models for te reo, ensuring a language too small for commercial markets remains under the control of its native speakers.
• Global Finance: PwC has fine-tuned an open model specifically for financial linguistics, serving hundreds of clients on private hardware without paying per-token fees.
• Humanitarian Aid: In Lausanne, a collaboration with the Red Cross has produced a medical model aligned with humanitarian ethics, now heading toward clinical trials in Tanzania and Switzerland.
• Agriculture: In East Africa, farmers are using offline, on-device models to identify cassava diseases in regions devoid of cloud connectivity.
• Public Infrastructure: A Swiss consortium leveraged public supercomputers to train a national model, releasing the weights, data, and training code for all to see.

"Mozilla was born because we refused to let a single entity own the gateway to the internet. We apply that same philosophy here: the future must be defined by competition and interoperability. We envision a diverse ecosystem of models and standardized interfaces that allow users to switch vendors without friction."

The era of renting intelligence is the only way. $\rightarrow$ The era of owning intelligence is here.

---

01 | The Current Landscape of Open-Source AI

The Closing Capability Gap

The distance between open-weight models and the "frontier" closed models has shrunk dramatically, though it remains uneven.

$\text{Capability Gap} \approx 3.3\% \text{ (as of March 2026)}$

While closed models still hold an edge in complex reasoning and multimodality, open models have reached parity in coding, general knowledge, and instruction-following.

The Collapse of Intelligence Costs

The cost of running GPT-4 class inference has plummeted faster than the historical price drops of PC compute or dotcom-era bandwidth.

---

02 | Market Adoption & Token Volume

Open weights are no longer just for hobbyists; they are dominating actual production workloads.

The "Token Win"

According to OpenRouter data, the volume of tokens routed through open-weight models has surged. By mid-2026, the top five highest-volume models were all open weights.

# Top Volume Models (OpenRouter Leaderboard)
top_models = [
    "Open-Weight-Model-A", 
    "Open-Weight-Model-B", 
    "Open-Weight-Model-C", 
    "Open-Weight-Model-D", 
    "Open-Weight-Model-E"
]
# Closed models like Claude follow behind in volume

Geopolitical Distribution

There is a stark divide in who is building and using these models. Chinese-built open models are seeing massive adoption compared to US-built ones.

• Chinese-built models: $\approx 18\text{T}$ weekly tokens.
• US-built models: $\approx 5.5\text{T}$ weekly tokens.
• Ratio: $> 3:1$

---

03 | Developer Trends & The Production Hurdle

While developers love open models, moving them into a production environment remains a challenge.

Adoption vs. Production

• Adoption: $79\%$ of developers use open models; $71\%$ use closed models.
• The Overlap: Most developers don't choose one; they use both.

The "Production Gap"

Despite high adoption, open models struggle to reach the final production stage compared to closed alternatives.

Analysis: This gap is not caused by a lack of capability, but by a lack of operational tooling. Enterprises can simply "buy" their way into a closed deployment, whereas open deployments require more manual infrastructure.

Why Teams Churn (The Operational Blockers)

The primary reasons developers abandon open models are consistent across all global regions:

• Infrastructure Costs: High GPU/hosting overhead.
• Security & Compliance: Difficulty in auditing and securing self-hosted stacks.
• Maintenance: The burden of updating and patching models.
• Deployment Complexity: Lack of "one-click" production pipelines.

Regional Nuance:

• South Asia: Heavily concerned with security and support.
• North America & Greater China: Only regions where more than $15\%$ of churn is attributed to specific performance gaps.
• South America & Western Europe: The only regions where closed-model adoption actually exceeds open-model adoption.