![\"Adina](\"https:\/\/cdn-avatars.huggingface.co\/v1\/production\/uploads\/63a369d98c0c89dcae3b8329\/AiH2zjy1cnt9OADAAZMLD.jpeg\"\/)
<\/span> <\/span> <\/div>\n![\"Irene](\"https:\/\/cdn-avatars.huggingface.co\/v1\/production\/uploads\/62543749b777cd32720675c2\/EF_KRZO4hTo8TWXOtvc-n.png\"\/)
<\/span> <\/span> <\/div>\n<\/div><\/div>\n<\/div>\n<\/div>\nWithin the AI world, time feels compressed. Just a few months after our spring report in our biannual evaluation labored by the ecosystem, there are fairly a couple of findings that now we have noticed till this summer time. This report lays out these observations from January to August 2026 and presents the info behind every one.<\/p>\n
![\"Cumulative](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/dataset-growth.png\")
<\/p>\n
Fashions and datasets on HF hub are rising every day. Public mannequin repositories grew from 2.43 to 2.96 million over the interval, datasets from 711,000 to 1 million, Areas from 1.00 to 1.44 million. The distribution beneath stays excessive, roughly 85.6% of fashions have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads. Every little thing beneath occurs inside that form.<\/p>\n
1. The frontier is shifting quick<\/p>\n
There was once a transparent development path: labs would begin by releasing smaller fashions and step by step work their approach towards the highest finish of the size. In 2026, a number of Chinese language labs skipped this development fully.<\/p>\n
![\"Largest](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/frontier-ceiling-by-country.png\")
<\/p>\n
In virtually each month of 2026, the most important and most performant open mannequin from a Chinese language lab was bigger than any mannequin an American lab launched. China’s month-to-month ceiling ran between 754B and a couple of.78 trillion parameters; U.S. fashions stayed beneath 130B in 5 of seven months, the exception being NVIDIA’s Nemotron 3 Extremely at 561B in Might and June, and Inkling from Considering Machines Lab.<\/p>\n
![\"Every](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/lab-size-strategy.png\")
<\/p>\n
The chart splits the labs into two camps. Moonshot, MiniMax, Xiaomi and Z.ai publish virtually nothing beneath 70B, so a developer’s first encounter with them is a mannequin too giant to run on something they personal. Tencent and Alibaba Qwen cowl the entire vary as an alternative, from beneath 1B upward.<\/p>\n
Two issues made the primary camp attainable. Constructing giant stopped being a differentiator. Xiaomi, Ant Group and Meituan each cleared a trillion parameters this yr, and neither was a family title in open weights twelve months in the past. And a lab not has to ship a small mannequin to be reachable, as a result of the group’s quantization layer will make a big one runnable inside days, a dependency we return to beneath.<\/p>\n
That leaves the scale profile as an announcement of intent moderately than of functionality. A frontier solely portfolio stakes every little thing on benchmark place and API demand. A full spectrum portfolio is a bid to be the household builders standardise on. Each are rational, they’re taking part in for various prizes.<\/p>\n
The US will not be absent from open supply.<\/p>\n
![\"New](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/new-homegrown-models.png\")
<\/p>\n
The 2 organizations publishing probably the most new open fashions this yr are additionally the businesses making the {hardware}: AMD and NVIDIA. Every launched greater than 200 new mannequin repositories, far forward of the remainder of the sphere, with LiquidAI rating third at round 100. {Hardware} distributors have realized that open fashions are a approach to promote chips: a mannequin optimized in your {hardware} and freely out there is the clearest proof that the {hardware} works.<\/p>\n
When smaller fashions and embedding fashions are included, the place Google, Microsoft, IBM Granite, and OpenAI\u2019s older imaginative and prescient and speech fashions generate a whole bunch of thousands and thousands of downloads yearly, U.S. open supply AI is rising.<\/p>\n
Extra {hardware} and infrastructure organizations comparable to NVIDIA are coaching and open-weighting aggressive fashions. NVIDIA’s Nemotron mannequin household boasts excessive efficiency. Lengthy-time leaders comparable to Meta reignite open roots with Meta’s Muse Glimmer.<\/p>\n
On the frontier scale, some U.S. mannequin releases above 100B parameters this yr are constructed on prime of Chinese language fashions or leverage artifacts from Chinese language labs, comparable to Considering Machines\u2019 Inkling (952B). Main unique American fashions embody NVIDIA\u2019s Nemotron 3 Extremely (561B), Nemotron 3 Tremendous (124B), and Arcee AI\u2019s Trinity-Massive (399B).<\/p>\n
AMD contributed many conversions. This work is necessary: it allows trillion-parameter fashions to run effectively on U.S. {hardware}. This represents a distribution and optimization layer.<\/p>\n
In the meantime, Chinese language open fashions are more and more optimized for home chips in China, the identical competitors in reverse, the place fashions are designed round particular {hardware} ecosystems.<\/p>\n
\n
\t We took the highest 25 mannequin repositories by downloads amassed this yr and the highest 25 by likes. Precisely one repository seems in each lists.<\/p>\n We counted downloads contained in the window moderately than lifetime, so nothing is credited for merely having existed longer, and controlling for age makes the break up sharper. Not one mannequin printed in 2026 reaches the obtain prime 25, whereas 13 of the twenty-five date from 2022. all-MiniLM-L6-v2 was pulled 1.55 billion instances in seven months in opposition to 5,156 likes; Kimi-K3 was pulled about 60 instances per prefer it obtained.<\/p>\n The 2 numbers report completely different acts. A like says a launch issues, and goes to frontier fashions within the weeks after they ship. A obtain says one thing is wired right into a pipeline that runs on a schedule, and accrues to small, secure fashions over years. Likes are the fitting instrument for studying what the sphere is happy about, downloads for studying what it presently is determined by. Treating both as a proxy for the opposite is the most typical mistake we see in protection of the Hub, together with our personal earlier work. The identical break up seems on the stage of the writer.<\/p>\n China’s frontier labs are the one accounts on the Hub the place the heavy band carries the amount. Successfully all of MiniMax’s 2026 downloads are of fashions above 70B, together with 88% of Moonshot’s, 55% of DeepSeek’s and 39% of Z.ai’s. No giant American account seems to be like this: Google, Microsoft and IBM Granite report basically none of their 2026 downloads above 70B, and NVIDIA and Meta solely 14% and 9%.<\/p>\n The distinction turns into clearer in complete downloads. Moonshot\u2019s frontier-only portfolio recorded 37M downloads over the yr, whereas Qwen\u2019s broader launch technique throughout mannequin sizes reached 2,045M (throughout repositories with declared parameter counts, 2,061M together with all repositories) , about 55 instances extra. The continued enlargement of the household, from the two.4T-parameter Qwen 3.8 Max to smaller variants comparable to 27B, reveals the identical give attention to protection throughout completely different use circumstances.<\/p>\n Time additionally performs an necessary function. Most fashions expertise a pointy decline in utilization after launch, adopted by an extended tail of regular exercise. A mannequin\u2019s adoption is essentially decided inside its first few months.<\/p>\n This helps clarify why right now\u2019s obtain quantity is usually pushed not by the most recent releases, however by a smaller group of fashions which have turn into established infrastructure over time.<\/p>\n \t If frontier fashions had been a licensing enterprise, you’ll count on the most important releases to hold the tightest phrases. Nevertheless, the info beneath reveals a special story.<\/p>\n Of 178 Chinese language releases above 20B parameters this yr, 59% carry Apache 2.0 and 22% carry MIT, and most carry a non-commercial restriction. Nevertheless, in the previous few weeks, we began to see a change on this development for the actually giant fashions, with Kimi K3 and Qwen3.8 beginning to embody some non-commercial restrictions and income share necessities to their licenses<\/p>\n DeepSeek and Z.ai ship fashions between 700 billion and 1.65 trillion parameters beneath plain MIT. Chinese language labs license their largest fashions about as permissively as their smallest, and extra permissively than American labs license theirs: on the American aspect of the identical dimension band, 29% is Apache or MIT, 41% sits beneath customized phrases and 30% declares nothing in any respect.<\/p>\n No matter these releases are for, it isn’t licence income. The weights are given away on probably the most permissive phrases out there. The return has to come back from someplace else: API and cloud enterprise, {hardware} and platform positioning, or the ecosystem place itself. As an example, the valuations of Z.ai and Kimi level to an efficient open supply technique, getting traction and development alternatives in the neighborhood. Going ahead, nevertheless, the business is more likely to shift towards clearer monetization paths from open-source adoption.<\/p>\n \t A mannequin\u2019s ecosystem place will not be outlined solely by its personal releases, however by how a lot the group builds on prime of it. As talked about above, Qwen is one exception which is getting consideration and adoption.<\/p>\n Knowledge from Hugging Face<\/p>\n By this measure, Qwen has turn into one of many largest foundations within the open mannequin ecosystem. Qwen-based fashions now account for 151,448 derivatives on the Hub, 2.6\u00d7 Meta\u2019s complete footprint and 4.7\u00d7 the Llama repositories particularly. Google follows with 82,506 derivatives. The third-largest supply is Unsloth, a group account publishing quantized and fine-tuning-ready builds, lots of which additional prolong the Qwen ecosystem.<\/p>\n Qwen derivatives have elevated at roughly 180\u2013210 new repositories per day all through the primary seven months of 2026, exhibiting that adoption will not be pushed solely by particular person launches. Qwen has turn into a part of the default workflow for builders deciding what fashions to fine-tune and deploy.<\/p>\n A number of components contributed to this place. First, consistency. Qwen has maintained an everyday launch cadence, repeatedly updating its mannequin household moderately than counting on occasional flagship releases. Second, protection. It publishes fashions throughout a variety of sizes and use circumstances, permitting builders to remain throughout the similar ecosystem whether or not they want a small native mannequin or a bigger deployment mannequin. Third, openness. Apache 2.0 licensing reduces friction for modification, redistribution, and industrial use.<\/p>\n These components reinforce one another. A broad mannequin household attracts extra builders; extra builders create extra derivatives; and people derivatives make the ecosystem extra engaging to future customers.<\/p>\n This place was constructed largely by the group. The 151,448 derivatives symbolize downstream work created by different builders, not releases produced by Qwen itself. Even among the many 28,531 GGUF conversions of Qwen fashions on the Hub, Qwen printed solely 54.<\/p>\n \t Amongst fashions that declare a parameter depend, these beneath 1B take 83% of all-time downloads and every little thing above 100B takes 1%. Limiting to downloads amassed in 2026 modifications nothing: 3% of the amount goes to fashions above 70B. That is the March discovering that has held up most cleanly, for a similar cause as earlier than, small fashions are the one ones that run on the {hardware} most builders even have.<\/p>\n So how does a trillion-parameter mannequin attain anybody in any respect? Via llama.cpp.<\/p>\n In February the ggml group joined Hugging Face, with the venture remaining absolutely open-source, community-governed and in the identical technical route. What modified is that crucial venture in native inference now has sturdy assets behind it.<\/p>\n The ceiling moved with llama.cpp. The July snapshot carries GGUF builds of DeepSeek-V4-Flash at roughly 284B parameters and Kimi-K3 at roughly 2.8 trillion. Native inference used to imply an 8B mannequin on a laptop computer. It now means a trillion-parameter mixture-of-experts unfold throughout a couple of shopper machines, which is the choice route the frontier didn’t have a yr in the past, and the explanation a frontier-first launch technique is viable in any respect.<\/p>\n And that route runs on Qwen: 39.6 million GGUF downloads a month, practically twice Gemma’s 20.8 million and greater than 5 instances Llama’s 7.5 million. The Llama hole will not be a provide drawback, Llama-derived GGUF repositories barely outnumber Qwen’s. Identical shelf area, a fifth of the visitors.<\/p>\n Mannequin repositories grew 21.5% over these seven months. A number of issues round them grew a number of instances quicker.<\/p>\n Repositories declaring the gguf library rose 464%, lerobot 194% and Apple’s mlx148%, in opposition to 16% for transformers and peft and 21% for diffusers. The modelling core is rising at roughly the platform common. The layer that decides the place a mannequin can bodily run native inference codecs, Apple silicon, robotic management stacks, is rising three to seven instances quicker than that.<\/p>\n Throughout the ten largest mannequin households, the labs behind these fashions publish only a few official GGUF conversions. But GGUF variations are sometimes those utilized by builders operating fashions regionally. Offering an official conversion at launch, documenting quantization selections, and signing the artifacts would require restricted extra effort. Reasonably than sustaining this workflow internally, labs may collaborate with current ecosystem contributors comparable to Unsloth. Doing so would cut the hole between the weights examined by mannequin creators and the variations adopted by the broader group.<\/p>\n \t We couldn’t have written this part in March, as a result of the instrument didn’t exist. The agent-usage dataset, printed in July, information the agent\/ token that coding brokers ship after they name the Hub by huggingface_hub or the hf CLI \u2014 trying to find fashions, pushing datasets, operating Jobs, creating Areas. For the primary time we are able to see how a lot agent visitors the Hub receives and which harnesses it comes from.<\/p>\n The second discovering is the unregistered row. Almost 1 \/ 4 of agent-tagged visitors in July got here from harnesses not but named within the dataset, and in Might that determine was 59.8%. Between April and July greater than a dozen new shopper identifiers appeared. New entrants are arriving quicker than any registry can title them \u2014 which is itself the discovering.<\/p>\n We spent a lot of the yr constructing for this reader moderately than just for human browsers. Papers started serving machine-readable Markdown in March. April introduced agent traces as a first-class dataset sort and an brokers.md endpoint on each Gradio Area, so an agent can learn a Area’s API and name it straight. July introduced the hf_fs software on our MCP server, exposing repositories, storage, docs and papers by a single interface in simply over a thousand tokens, alongside attachable sandboxes for safe execution. The identical consolidation occurred on the protocol layer, with MCP shifting into the Linux Basis’s Agentic AI Basis.<\/p>\n Then, in July, an agent stopped being a reader and have become an intruder. What seems to be the primary documented case of an autonomous agent operating a sustained intrusion by itself initiative occurred to us. Whereas our group tried to make use of frontier closed fashions to research the captured assault code, their security guardrails declined the work. The evaluation was accomplished ultimately on a quantized open mannequin GLM-5.2 operating on our personal infrastructure. We printed a disclosure and a full technical timeline.<\/p>\n \t In comparison with the spring report, the geographical rebalancing of energy continues to speed up. Whereas U.S. open supply fashions proceed to be aggressive, the race between a number of Chinese language frontier mannequin labs attracts sturdy consideration. Many likes on these frontier fashions level to what excites the group probably the most, and development alternative for firms leveraging the eye for valuations.<\/p>\n Nevertheless, the AI race will not be solely sprints, but additionally a marathon; instruments like llama.cpp helps deploying the massive fashions regionally, however a broad mannequin household and its adoption remains to be the important thing, to construct a optimistic suggestions loop between builders, writer and future customers. Fashions to be embedded within the infrastructure and being a part of the ecosystem, might result in a commercially sound exit on the finish of the tunnel.<\/p>\n In the long run, with brokers being the number one person on HF Hub for the primary time, the subsequent report might look very completely different.<\/p>\n In AI, a couple of months can reshape the ecosystem.<\/p>\n
\n\t\t2. Consideration \u2260 Adoption
\n\t<\/span>
\n<\/h2>\n![\"Attention](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/attention-vs-usage.png\")
<\/p>\n![\"Who](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/downloads-by-lab-and-model-size.png\")
<\/p>\n\n
\n\t\t3. Open weights shift the place worth accumulates
\n\t<\/span>
\n<\/h2>\n![\"The](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/model-licenses-by-region-and-size.png\")
<\/p>\n\n
\n\t\t4. Qwen has turn into the group’s base mannequin
\n\t<\/span>
\n<\/h2>\n![\"Derivatives](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/derivatives-by-organization.png\")
<\/p>\n\n
\n\t\t5. Small fashions stay the sensible layer
\n\t<\/span>
\n<\/h2>\n![\"Downloads](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/downloads-by-model-size.png\")
<\/p>\n![\"llama.cpp](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/llama-cpp-hub-growth.png\")
<\/p>\n![\"What](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/local-downloads-by-model-family.png\")
<\/p>\n![\"The](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/runtime-layer-growth.png\")
<\/p>\n\n
\n\t\t6. Brokers are the brand new person
\n\t<\/span>
\n<\/h2>\n![\"Agents](\"https:\/\/huggingface.co\/datasets\/huggingface\/documentation-images\/resolve\/main\/blog\/state-of-open-models-summer-2026\/agent-hub-traffic.png\")
\nClaude Code led July with 44.4%, however a single month conceals the actual discovering: it held 67.8% in April and 6.4% in Might, whereas Codex climbed steadily from 10.4% to twenty.8%. It is a market with no incumbent, the place one launch or one modified default can transfer half the visitors in a month.<\/p>\n\n
\n\t\tWanting ahead
\n\t<\/span>
\n<\/h2>\n\n