Are Closed LLMs Losing Their Lead? Open-Weight Models Are Moving Closer to the Frontier
Inkling, Kimi K3, and Qwen3.8 changed the open-weight LLM landscape within just a few weeks, raising important questions about the future of the AI race.
As we entered the summer of 2026, the strongest players in the large language model race still looked very familiar: OpenAI, Anthropic, and other closed-model providers. Anthropic released Claude Fable 5 on June 9, 2026, while OpenAI made the GPT-5.6 family generally available on July 9 following a limited preview that began on June 26.
Both models were introduced with strong reasoning, coding, and agentic capabilities. Their advanced capabilities and potential security risks, particularly in cybersecurity, also brought them into broader discussions involving the U.S. government.
Then, within only a few weeks, three major developments changed the picture on the open-weight side.
🇺🇸 July 15, 2026 — Inkling: Thinking Machines Lab released Inkling, an open-weight model with 975 billion total parameters and 41 billion active parameters per token.
🇨🇳 July 16, 2026 — Kimi K3: Moonshot AI announced Kimi K3, a 2.8-trillion-parameter model. Its full model weights were released on July 27.
🇨🇳 July 19, 2026 — Qwen3.8 preview: Alibaba introduced Qwen3.8-Max-Preview in the 2.4-trillion-parameter class. Qwen3.8-Max was formally introduced on August 3, followed by the open-weight Qwen3.8-2.4T-A95B model on August 12.
🇺🇸 Inkling: A major open-weight move from the United States
The significance of Inkling is not simply that it has 975 billion parameters. The model uses a Mixture-of-Experts architecture with around 41 billion active parameters per token and can reason across text, images, and audio.
It also supports agentic coding, tool use, and context windows of up to one million tokens.
What makes Inkling particularly interesting, in my view, is something else. While Chinese companies had become far more visible in frontier open-weight model development, this time a powerful and genuinely open-weight model came from a new U.S.-based AI laboratory.
🇨🇳 Kimi K3: Open-weight models approach the 3-trillion-parameter class
Immediately after Inkling, Moonshot AI introduced Kimi K3, pushing open-weight model scale into a new range.
The model contains 2.8 trillion parameters. Its MoE architecture includes 896 experts, while only 16 are activated for each token. This allows the total model size to become extremely large without requiring every parameter to participate in every inference step.
Kimi K3 also attracted attention for more than its size. Around the time of its release, it ranked first on Frontend Code Arena, ahead of Claude Fable 5.
That does not mean Kimi K3 is now simply “better than GPT-5.6 or Fable 5.” Closed frontier models still hold an advantage in many broad evaluations. But the fact that an open-weight model can outperform the strongest closed models in selected coding and agentic tasks is a substantial change compared with the situation only a few years ago.
🇨🇳 Qwen3.8: Alibaba also moves into trillion-scale open models
Alibaba’s Qwen3.8 move showed that these developments were not isolated events.
Qwen3.8-Max, previewed in July, reached its formal 2.4-trillion-parameter release in August. The Qwen team then released the weights of Qwen3.8-2.4T-A95B as well.
As a result, within a short period, the open-weight ecosystem saw new models at 975 billion, 2.4 trillion, and 2.8 trillion parameters.
A 2.8-trillion-parameter model is not automatically better than a 500-billion-parameter model. Training data quality, architecture, post-training, reinforcement learning, tool use, inference infrastructure, and active parameter count can matter just as much as total model size. The real shift is that open-weight models are beginning to operate in the same capability range as frontier systems.
🔓 Are open-weight models still six or seven months behind?
For several years, there was a fairly reasonable rule of thumb: the strongest capabilities first appeared in closed systems from companies such as OpenAI, Anthropic, or Google, while comparable open models followed months later.
That assumption now needs to be reconsidered.
Models such as Kimi K3 can enter the same performance region as GPT-5.6 and Claude Fable 5 on selected tasks. The Qwen family continues to improve rapidly in coding and agentic workloads. Inkling shows that a U.S.-based open-weight frontier model is also possible.
Open models have not yet surpassed closed models across every dimension. But in many areas, the delay no longer looks like “wait for the next model generation.”
To me, this is the more important change.
🇨🇳 Are Chinese companies catching up with U.S. frontier labs?
Recent developments make this question increasingly difficult to dismiss.
Chinese companies are no longer producing only smaller or cheaper alternatives. With Moonshot AI and Alibaba, we are now seeing models at true frontier scale, and some of them are being released with their weights.
This is particularly significant because Chinese companies are operating under restrictions affecting access to some of the most advanced U.S. AI chips. Despite this, they continue to scale through work on model architecture, sparsity, attention mechanisms, quantization, and distributed inference.
The competition may therefore increasingly shift from “Which company has the best model?” to “Which national ecosystem can still develop frontier models?”
🔐 Will Chinese companies eventually close their strongest models?
This may become one of the most interesting questions for the future of the open-weight ecosystem.
Chinese model developers have benefited significantly from open-weight strategies. Their models have spread quickly across the world, developer communities have formed around them, inference platforms have added support, and the companies themselves have gained global visibility.
But would a company follow the same strategy if it genuinely believed it had the most capable model in the world?
Or would it become commercially more attractive to keep the strongest model behind an API while releasing the previous generation as open weights?
The next few model generations may provide an answer.
🌍 Where are the countries outside the U.S. and China?
This may be one of the least discussed aspects of the current AI race.
Two ecosystems now stand out very clearly in frontier model development: the United States and China.
Europe has strong research institutions and important AI companies. Japan and South Korea have advanced semiconductor industries. India has a very large software and engineering talent base. Yet at the trillion-parameter frontier, the competition is increasingly concentrated in the U.S. and China.
The reason is that this race cannot be run with good researchers alone. It requires tens of thousands of accelerators, data-center infrastructure, energy, capital, training data, distributed-systems engineering, and a very large research ecosystem.
One of the strategic questions for the coming years will therefore be whether other countries attempt to build their own frontier-model infrastructure or become increasingly dependent on models developed in the U.S. and China.
🔹 Can Chinese companies fully catch up with U.S. frontier-model companies technologically?
🔹 Will the performance gap between open-weight and closed models disappear within the next few generations?
🔹 To what extent will independent evaluations confirm the superiority suggested by selected benchmarks from companies such as OpenAI and Anthropic?
🔹 Will Chinese companies continue their open-weight strategies even for their most capable models?
🔹 Can countries outside the U.S. and China re-enter the frontier-model race?
🚀 Is a new era beginning for open-weight LLMs?
It is still too early to say that open-weight models have completely overtaken closed systems.
But something else can now be said with much more confidence: open-weight models are no longer sitting at the edge of the frontier AI race. They are moving closer to its center.
The rapid appearance of Inkling, Kimi K3, and Qwen3.8 is one of the clearest signs of this shift.
If the performance gap between open and closed models continues to narrow, the consequences will extend far beyond benchmark tables. It will affect how AI infrastructure is built, which model providers companies depend on, how countries think about AI sovereignty, and even the future of today’s API-centered business models.
What do you think? When will open-weight models catch closed models in overall performance? Or will the most capable models always remain closed?
Murat Karakaya Academy
#MuratKarakayaAkademi #ArtificialIntelligence #LLM #OpenWeights #OpenSourceAI #KimiK3 #Qwen38 #Inkling #GPT56 #ClaudeFable5 #AgenticAI #GenerativeAI
