Mozilla State of Open Source AI 2026 Report Highlights

Open‑weight models have reached parity for most workloads

Open‑weight models now match closed models on coding, instruction‑following and general knowledge, with a 0.5% capability gap in early 2025 that widened to 3.3% by March 2026 mainly on reasoning, long‑context retrieval and agentic tasks. The five highest‑volume models on OpenRouter are all open, and open models route >50% of production tokens by mid‑2026.

"Open weights are no longer a compromise. They are where the work happens: a majority of production tokens now route through them" – Mozilla CTO Raffi Krikorian

Token‑volume vs. revenue

OpenRouter data (Nov 2024‑Nov 2025) shows open‑weight token share grew from negligible to ≈33% by late 2025 and a majority by mid‑2026.
Closed models still dominate request count, but open models lead in token volume, especially for coding and agentic workloads.
In May‑Sep 2025 closed models captured ≈96% of revenue while handling ≈80% of usage, reflecting a ≈6× price premium for comparable capability.

Inference cost has collapsed dramatically

GPT‑4‑class inference fell from $20 → $0.40 per 1M tokens in 36 months – a 50× reduction.
Stanford HAI AI Index 2025 and MIT studies confirm a 5‑10× yearly decay at the frontier, outpacing hardware cost curves.
The price drop makes open models economically attractive for large‑scale production.

Adoption is high but production stalls on operational gaps

Mozilla/SlashData 2026 developer survey: 79% of developers adding AI use open models, 71% use closed models; 50% use both.
Only 51% of open‑model teams reach production versus 63% for closed‑model teams.
The biggest churn drivers are performance, integration, and maintenance – operational, not capability issues.

Challenge	% of churned developers
Performance not good enough	+12pp
Integration into existing systems	+11pp
Ongoing maintenance & updates	+10pp
Insufficient documentation	+8pp
Deployment, hosting, scaling	+8pp

Regional adoption patterns

Greater China and East Asia lead with 89% open‑model adoption.
Western Europe (70%) and South America (66%) still favor closed models.
Enterprise size does not close the operational gap: open‑model production rises only from 53% → 57% when moving from small to enterprise teams, while closed‑model production climbs from 54% → 73%.

The open‑source AI stack: strong capabilities, weak operations

The nine‑layer stack (48 components) scores ≥4.0 on capability but <2.5 on standardisation and enterprise readiness – the “operational gap”.

Market momentum and business models

Venture‑backed open‑weight companies are now multi‑hundred‑million‑dollar businesses: e.g., DeepSeek ($7.4 B funding, >$220 M ARR), Mistral AI (~$400 M ARR), Databricks ($5.4 B run‑rate).
Five proven revenue models: hosted inference, enterprise platforms, on‑prem licences, fine‑tuning services, and harness tooling.
Large enterprises are fleeing metered pricing: Microsoft cancelled Claude Code licences after token‑billing exhausted its AI budget; Uber capped AI spend after blowing through its coding budget; Stripe cut inference costs 73 % by self‑hosting open models on vLLM.

Sovereignty and policy drivers

Over 70 national AI strategies emphasise ownership of AI stack layers.
A June 2026 export‑control incident forced Anthropic to cut access for foreign nationals, demonstrating the exit‑risk of closed APIs.
Moving a petabyte out of AWS S3 costs $90‑120 k, prompting 80% of enterprises to repatriate workloads.
China’s “AI Plus” Initiative and Five‑Year Plan deliberately push open‑weight releases to hedge against semiconductor export controls.
Hugging Face downloads (Mar 2026): Qwen (942 M) vs. Llama (476 M), showing Chinese open models dominate raw download volume.

The agentic harness: the next competitive layer

The harness (orchestration loop, tools, memory, permission model) is where capability is engineered today.
Benchmarks (Terminal‑Bench 2.0 & 2.1) show third‑party harnesses can outperform closed‑model labs on the same weights, but labs quickly pull the harness in‑house, creating a lock‑in effect.
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; existing specs (MCP, A2A) stop at authentication, not authorization.
Emerging meta‑harnesses (e.g., Databricks Omnigent) enforce stateful write policies across harnesses, offering a potential solution.

Where closed models still lead

Integrated harnesses that are tightly tuned to proprietary weights.
Long‑context retrieval (e.g., Gemini 3 at 89 % vs. DeepSeek V4‑Pro at 41 %).
Turnkey compliance (SOC 2, HIPAA) baked into closed offerings.
Legal accountability – a clear contractable counter‑party.

Five strategic bets for the open ecosystem

Bet	What to own	Why it matters
1. Build an open harness	Co‑design with open weights; deliver general‑purpose or vertical‑specific orchestration.	Prevents closed labs from welding model + harness into a single rented product.
2. Own the memory layer	Portable, append‑only context stores behind your firewall.	When model cost → 0, memory becomes the lasting asset.
3. Define portable permission	Open standard for write‑side policies across MCP, A2A, and tool calls.	Locks in governance before closed platforms set the de‑facto rule.
4. Break the meter	Run a second‑source model in‑house for predictable loads; self‑host when token pricing spikes.	Metered pricing is a rent‑trap; ownership converts OPEX to CAPEX.
5. Keep the open default plural	Foster multiple open‑weight suppliers (e.g., China, Europe, US) to avoid a single‑source commons.	Prevents the “open‑default” from becoming a new monopoly.

Watchlist signals

Capability gap widening beyond 3 % → risk of open models falling behind.
Harness dominance expanding (lab‑owned vs. independent) → lock‑in risk.
Funding slowdown for open‑weight companies → could stall ecosystem growth.
Safety incidents (e.g., high‑CVSS authorization failures) → could trigger regulation that harms open models.

Community reactions (selected HN comments)

"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. … 4.19T tokens processed today, almost 5× 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

"There’s 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

All figures and citations are taken directly from Mozilla’s State of Open Source AI 2026 report and the linked sources.