Mozilla report: Chinese open AI models close to parity with US frontier models
Performance gap has narrowed to 4.4 months; open models cost 70% less but lack compliance packaging.
What to know
- Performance gap between Chinese open models and US closed frontier models has narrowed to just 4.4 months—meaning an open model released today will match current frontier performance in that timeframe.
- Open models cost 70% less than frontier alternatives while achieving near-identical benchmark scores, making them the economical default for most organizational workloads.
- The remaining value of closed models lies in specialized high-complexity tasks (8-12 hour reasoning jobs); most organizations lack internal expertise to fully optimize open model deployment, giving closed providers a compliance and support advantage.
- Real-world adoption is accelerating: companies like DoorDash now use open models as default, reserving expensive closed models only for expert-level work.
Open models are now economically dominant for routine work; closed models face serious margin pressure.
-
“interesting results that are not shocking from this Mozilla study on open weight vs. frontier AI models”
wd.teleri.net · Bluesky ↗
Closed model vendors retain defensible positions through compliance, support, and superior UX rather than pure capability.
-
“i don't think it's as simple as "closed frontier model companies are cooked" but i guess their one big hope is that no one ever makes better harnesses and compliance packaging for open models ever?”
lauren.rotatingsandwiches.com · Bluesky ↗
“We see the decision to pay for closed [models] as workload-specific rather than organization-specific.”
Raffi Krikorian, CTO, Mozilla · Ars Technica ↗ · Sep 14
Mozilla Research organizationRaffi Krikorian Chief Technology Officer, MozillaAnthropic Closed frontier model providerMoonshot AI Open model providerDoorDash Enterprise adopter
How it unfolded 2 developments, newest first · click a bar or a number to jump articlesposts
-
2
Companies adopt open models for routine tasks based on cost-performance ratio
Real-world deployment patterns emerge showing organizations like DoorDash using Kimi for routine work while reserving Fable for complex tasks. The cost advantage and narrowing capability gap are accelerating adoption of open models since Mozilla's July report.
“If the open frontier can handle a seven-hour job, the closed frontier can handle a 12-hour one. In four months, the open model handles the 12-hour job, and the closed one handles something around 20.”
— Raffi Krikorian -
L
arstechnica.com/ai/2026/09/e... i don't think it's as simple as "closed frontier model companies are cooked" but i guess their one big hope is that no one ever makes better harnesses and compliance packaging for open models ever?
2 more of the top 3 · 13 posts in this stretch
-
A
Exclusive: Open Chinese models close gap with Silicon Valley’s frontier AI models Ars previewed Mozilla’s report on how cheap open models caught up on capability. https:// arstechnica.com/ai/2026/09/exc lusive-open-chinese-models-close-gap-with-silicon-valleys-frontier-ai-models/?utm_brand=arstechnica&utm_social-type=owned&utm_source=mastodon&utm_m…
-
D
Second, it leaves the field open for the further development of Chinese "open weight" models that customers can run on their own modest, low-powered computers, which are presently lagging the US "frontier models" by a mere four months:
-
-
1
Mozilla finds performance gap between open and closed models has narrowed to 4.4 months
Mozilla's updated State of Open Source AI report shows Chinese open models like Kimi K3 achieving near-parity with US frontier models. Kimi K3 scores just three points behind Anthropic's Fable 5 on the Artificial Analysis Intelligence Index while costing only 30% as much. The report recommends organizations use open models as default for most work.
“[A Closed model] earns its premium in a few places: expert professional work, high-intensity retrieval, and long context. We see the decision to pay for closed [models] as workload-specific rather than organization-specific.”
— Raffi Krikorian -
background
Mozilla publishes inaugural open-source AI report — Mozilla releases its first State of Open Source AI report, establishing baseline measurements for comparing open-weights and closed frontier AI models.
Also covered reported alongside — the timeline has no entry for these yet
-
first by Mastodon, 12d ago · also Ars Technica
and 1 smaller piece
What people are saying 8 voices from 2 sites · best of 13 · verbatim
- Sep 16
-
P
I wonder how much this has to do with the AI executive handwringing lately... 🤔
-
P
But there's two giant problems with this: first, it is incredibly, fantastically illegal under antitrust law, because it represents a conspiracy among the dominant players to cease to compete with one another, and; second, it leaves the field open for the further development of Chinese "open weight" models that customers can run on their own…
-
K
The frontier Anthropic & OpenAI models are only ~4 months ahead of open models — at 5x the price. “Most organizations should ideally be using open models as the default for the majority of their work, according to [a report] from Mozilla…” [arstechnica.com]
- Sep 15
-
Open Chinese models close gap with Silicon Valley’s frontier AI models.
-
Related, from an @arstechnica.com article on open weight LLM models being close behind frontier models, in terms of development, while being far cheaper to run
-
K
A product built on stolen content that is itself trivially easy to steal? That's sort of hilarious.
-
W
interesting results that are not shocking from this Mozilla study on open weight vs. frontier AI models:
-
D
Open Chinese models close gap with Silicon Valley’s frontier AI models. arstechnica.com