Mozilla State of Open Source AI 2026 – Key Findings and Implications

Open‑weight models now dominate token volume while closed models still lead in revenue

Takeaway: By mid‑2026, open‑weight AI models route more than half of all production tokens on OpenRouter, yet closed‑model APIs capture about 96 % of model‑layer revenue because they remain pricier and more entrenched in enterprise tooling.

• Open‑weight models account for ~33 % of token volume in late 2025 and have grown to a majority by mid‑2026 (OpenRouter 100‑T token study). The five highest‑volume models on OpenRouter are all open weights (DeepSeek V4 Flash, MiMo‑V2.5, Hy3, MiniMax M3, Owl Alpha).
• Closed models still dominate request count and revenue: closed US providers lead on request volume, and closed models generated ~96 % of revenue despite representing only ~20 % of token usage (May‑Sep 2025 data).
• Inference cost for GPT‑4‑class models fell 50× in 36 months (from $20 to $0.40 per 1 M tokens), collapsing the price barrier for open models.

"Open‑weight AI has collapsed the capability gap while the price of intelligence has collapsed. 27× fall in GPT‑4‑class inference cost in 36 months" – Mozilla report.

Capability gap narrowed to a few percent, but reasoning still lags

Takeaway: Open models are at parity on coding, instruction‑following, and general knowledge, with an average 3.3 % gap on the Chatbot Arena benchmark; the remaining gap concentrates in reasoning, long‑context retrieval, and agentic tasks.

• Gap collapsed to 0.5 % by Aug 2024, briefly matched by DeepSeek‑R1, then reopened to 3.3 % by Mar 2026 as closed models pulled ahead on reasoning.
• Open models lead token volume for coding and agentic workloads, where most production demand lies.

"The question is no longer whether open models are good enough. It's what you need for your workload." – Mozilla report.

Adoption is high but production conversion lags

Takeaway: 79 % of developers adding AI functionality use open models, yet only 51 % of open‑model projects reach production, compared with 63 % for closed models.

• Survey (Mozilla/SlashData 2026, n=1,410) shows developers often run both open and closed models (50 % use both, 29 % open‑only, 21 % closed‑only).
• The primary churn reasons are operational: performance perception (+12 pp), integration (+11 pp), maintenance (+10 pp), documentation (+8 pp), and deployment complexity (+8 pp). Security, privacy, and compute cost are not cited as major blockers.

"The biggest gaps (performance, integration, maintenance) are operational, not capability." – Mozilla survey.

Regional adoption patterns reveal sovereign interest

Takeaway: Open‑model adoption is strongest in Greater China and East Asia (≈ 89 % of developers), while Western Europe and South America still favor closed models.

• Chinese open‑weight models grew from <2 % of tokens (late 2024) to >45 % of weekly traffic on OpenRouter by Apr 2026, reaching 61 % among the ten most‑used models.
• National AI strategies (70+ countries) treat open weights as a sovereignty tool, enabling exit from vendor‑locked APIs.

"A provider can switch off a model. Nobody can switch off a copy already running on a machine you hold." – Mozilla report.

Business ecosystem shows multi‑billion‑dollar traction

Takeaway: Venture‑backed open‑source AI companies collectively raise tens of billions and generate hundreds of millions in ARR, proving a sustainable commercial model.

Company	HQ	Layer	Disclosed Funding	Valuation	ARR / Revenue Signal
Databricks	USA	Enterprise platform	—	—	$5.4 B run‑rate
DeepSeek	China	Frontier weights	$7.4 B	$50 B+	~$220 M ARR
Mistral AI	France	Weights + platform	$3.05 B	~$14 B	~$400 M ARR
Moonshot AI (Kimi)	China	Weights	$3.9 B	—	—
Cohere	Canada	Enterprise / on‑prem	$1.7 B	—	Command A+ open‑sourced May 2026

Five proven revenue models: hosted inference, enterprise platforms, on‑prem licensing, fine‑tuning services, and harness tooling.

Metered pricing drives enterprises to open models

Takeaway: High per‑token costs of closed APIs force large firms to switch to self‑hosted open models, delivering dramatic cost savings.

• Microsoft cancelled most Claude Code licenses after token billing exhausted its AI budget in months.
• Uber exhausted its 2026 AI coding budget in four months, then capped spend at $1,500 per tool per employee.
• Stripe cut inference costs 73 % by serving open models on vLLM, handling 50 M daily API calls with one‑third the GPU fleet.
• Estimated $24.8 B in annual savings from the price gap (≈ 6× cheaper per call for comparable capability).

The agentic harness is the next competitive frontier

Takeaway: While open weights are commoditizing, the agentic harness (orchestration, memory, permission layers) determines real‑world performance and lock‑in.

• Benchmarks (Terminal‑Bench 2.0 & 2.1) show a 21.8‑point gap between the best third‑party harness on Anthropic weights and the lab’s own harness; the gap later shrank to ~3 points after labs internalized the harness.
• No open model appears in the verified top tier of the official Terminal‑Bench 2.1 board, highlighting the current lack of a first‑party open harness.
• The write surface (actions that cause side‑effects) lacks a portable permission standard across MCP, A2A, and tool‑call frameworks, creating a security gap.
• Emerging meta‑harness solutions (e.g., Databricks Omnigent) enforce stateful policies—cost caps, human approvals, revocation—above individual harnesses.

"The unsolved permission problem is a write problem… emerging cross‑harness architectures enforce stateful, contextual policies." – Mozilla report.

Opportunities for builders

Takeaway: Five strategic bets can capture value without beating the frontier:

1. Build an open harness co‑designed with open weights, targeting verticals where closed labs have not yet integrated.
2. Own the memory layer (append‑only, portable formats) to retain context when models are swapped.
3. Define portable write permissions to become the de‑facto standard before closed platforms lock it down.
4. Break the meter by self‑hosting open models for predictable loads before token pricing spikes.
5. Promote a plural open default by supporting multiple open‑weight sources, preventing a single‑supplier commons.

Each bet has a time‑sensitive “clock”—the longer open layers stay fragmented, the more likely closed ecosystems will weld model and harness into a rented product.

Watchlist signals

Takeaway: Monitor four signal groups; reversal in any indicates a shift away from open dominance.

• Capability & adoption: If token share stalls while reasoning gap widens, open advantage erodes.
• Harness gap: Widening lead of lab‑owned harnesses or a closed permission spec becoming standard.
• Market structure: Failure of open‑lab ARR growth or a sovereign funding drop.
• Trust & safety: Major misuse events or regulatory moves to restrict open weights.

---

Community reactions on Hacker News

"Speculation: open models is what will kill Anthropic and OpenAI. … The harness is what takes these random and hallucinogenic models and make them into something deterministic and useful." – @babblingfish

"Exactly 4 months ago, the marketshare on openrouter was 60‑40 in favor of closed models. Now it’s 63‑37 in favor of open models… almost 5x growth in 4 months!" – @GodelNumbering

"I’m not ready to celebrate the victory of open models just yet considering all the good ones are built by private, VC‑funded companies. How long will they continue to be charitable?" – @paxys

"The design and layout made it harder to read than it needed to be. Regardless, the inference costs dropping almost 50× is really amazing to see." – @amanharshx

"I see a gap in the ecosystem: too few mature open source harnesses. I’d like a community‑led, BYOK, modular project where I can define, orchestrate, monitor and maintain agents." – @thih9

These comments echo the report’s core messages: rapid token‑volume growth for open models, operational challenges, and the critical need for open‑source harnesses and standards.