A Mozilla report on open weight Artificial Intelligence (AI) models draws the current panorama of this industry in which the “open” (models such as Kimi K3, Qwen or DeepSeek) has caught up with the closed (ChatGPT and Claude) in capacity and price, to the point that the best open model has been placed three points behind the closed leader with a price of 60 percent.
Mozilla has published version 1.1 of its report ‘The State of Open Source AI’ this September in which, apart from showing how the distance between closed and open models is reduced (those whose parameters are published and anyone can download and execute), it shows that open weights continue to fail in deployment.
The data provided by the report’s ‘benchmarks’ show that Kimi K3 is the open weight model that has questioned the frontier ones Fable 5 from Anthropic and Sol from OpenAI by reach number one in the LMArena rankings for the ‘frontend’ code, and heading six of the seven sections in this category in its premiere, in addition to general programming, following instructions and general knowledge.
Where it is practically tied with OpenAI’s Sol is in automated tasks (88.3 vs. 88.8), although it loses to Fable 5 in resolving real software incidents (81.2 vs. 86.6). And it loses to closedness in professional knowledge work, with Fable 5 taking its biggest advantage over K3, as in the handling very long texts without losing the thread and naturalness in conversation.
Contra Fable 5the best open model is only two points away, but costs 30 percent of what the closed one costs, which causes a series of economic implications, although deployment remains one of the biggest weaknesses of Chinese open weights.
‘TOOLING’, THE GREAT STILL
El ‘tooling’ (software tools surrounding the model) is the biggest problem faced by companies or developers who prefer open weights. According to the report, the open reaches production 12 points less than the closed and only captures 4 percent of the income.
With these two data it is easier to understand that, although 79 percent of developers use open models, only 51 percent have brought them into production. A figure that clashes with the 63 percent of those closed, who continue to win in this sense.
“This gap indicates that we are not facing a problem purely of quality of the model, but of absent infrastructure. The deployment rates of open models barely increase with the size of the company, which shows a lack of mature tools and support,” says SlashData Market Analyst, Álvaro Ruiz Cubero, regarding the Mozilla survey carried out with more than 950 developers, and in which he emphasizes that “buyers are prioritizing licensing conditions (31%) and ownership (26%), indicating a clear shift toward control and flexibility over raw capacity.”
This August was the first month in which an open model led OpenRouter (the platform that routes requests to multiple AI models) in number of requests, after 51 weeks of leadership from Google, until a DeepSeek model took it from it on August 3. Furthermore, the first four frontier models are closed, while the next four are open.
Besides, eight of the ten models with the highest volume of tokens are open pesosand precisely seven of them are made in China. In fact, China and East Asia lead the adoption of open AI with 89 percent, a figure well ahead of the West and which has become a national strategy.
ECONOMIC ISSUE
That companies look towards open weight models, especially when they are closer in performance, is mainly due to an economic issue. Uber, according to Bloomberg, exhausted its annual AI budget in four months by Claude Code token payment. Each closed model bills for each fragment of text processed, not for a flat rate, which increases costs with intensive use. Uber’s response was to cap it at $1,500 per month per employee.
The same thing happens with instability of the American frontier models, which for companies is a factor to monitor. Especially with the Fable 5 blackout episode, since, after its launch, it was inaccessible for 19 days due to an order from the United States Department of Commerce. An open weights model, on the other hand, runs on the company’s own machine and is always available.
Finally, in addition to the perspective of the economy and the availability of an AI model, the Mozilla report highlights that the battle in the industry is no longer fought in the model, but in el ‘harness’which is the ‘software’ layer that decides what an AI agent can see, remember and do.
In this sense, in May, independent tools were 21.8 points ahead of those from the laboratories themselves. But this new battlefield does not favor the open weights, but rather it is a new front in which the large laboratories counterattacked with their own internal ‘harness’ and, in eight weeks, reduced that advantage to three points. An example of those closed ‘harness’ is Claude Code of Anthropic.