Comment: OpenAI had been expected to live up to its name by releasing its first open-weights models since GPT-2 in this week. According to CEO Sam Altman, a safety review has delayed what could have been US’s first decent open model for the year. He wrote “While we trust the community will build great things with this model, once weights are out, they can’t be pulled back. This is new for us, and we want to get it right,” in a post at X.
This delay has left the US in an awkward position. Despite hundreds billions in GPUs being invested, the best open-model America has managed this year is Meta’s Llama 4. It received a less than stellar response and was marred by controversy. Just this week, it was reported that Meta had apparently taken its two-trillion-parameter Behemoth out behind the barn after it failed to live up to expectations.
A few other open models have been released by US companies. Microsoft released a version Phi-4 14B that was trained with reinforcement learning for reasoning functionality. IBM released a few tiny LLMs focusing on agentic workloads. Google released its multimodal Gemma3 Family, which had a maximum of 27 billion parameters. But these models are small fry compared to Meta’s 400-billion-parameter Llama 4 Maverick.
Among US companies, the real progress made in generative AI this year is locked away and only accessible through API calls to another’s servers.
China’s AI hot streak continues
While US model builders continue their best work behind closed door, China is doing so in the open. Nvidia CEO, Jensen Huang, likes to mention that half of the world’s AI researchers are Chinese, and it shows.
In 2025, DeepSeek became a household brand after the release of its R1 model . DeepSeek was a relatively unknown AI developer spun out of Chinese quant hedge fund High Flyer at that time.
The 671-billion-parameter LLM featured a novel mixture-of-experts (MoE) architecture that allowed it to run far faster and on fewer resources than even smaller LLMs like Llama 3.1 405B while replicating the reasoning functionality of OpenAI’s still fresh o1 model.
But more importantly, the model weights and technical documentation showing how it was done were made public. It should have been no surprise that Western developers began to replicate this process to give their models reasoning capabilities.
Since that time, Alibaba has released new reasoning and MoE model including QwQ. Qwen3-235BA22B and 30BA3B.
In June, Shanghai-based MiniMax released its 456-billion-parameter reasoning model called M1 under a permissive Apache 2.0 software license. The developer claims that a large context window with a million tokens and a new attention system help it keep track of them all.
Baidu opened sourced the Ernie family of MoE Models in the same month. The models range from 47 billion parameters up to 424 billion. Huawei also opened sourced its Pangu model trained on its own accelerators. However, this release was almost immediately overshadowed by allegations of fraud.
That brings us to July, when Moonshot AI, another Chinese AI dev, lifted the curtain on Kimi 2, a one-trillion-parameter MoE model they claim bests even the West’s most potent proprietary LLMs. Take those claims with a grain of salt, but the fact remains, the Chinese have developed a one-trillion-parameter open-weights model. The only US LLMs that come close to this are proprietary.
It is important to remember that all of this was achieved despite Uncle Sam’s crusade against the Chinese to prevent them from competing in the AI arena.
It’s not over yet.
Which brings us back to OpenAI and its open-weights model. Altman, the AI hype man, has only shared a few details about it in interviews with X and in Congressional hearings. Altman started the whole thing in February, when he asked his followers which OpenAI’s open source project they would prefer: an o3 mini-level model running on GPUs or their best smartphone LLM. The o3 mini-level LLM was chosen.
In June, OpenAI delayed its release for the first. Altman posted that the “research team did something unexpected and quite amazing, and we think it will be very very worth the wait but needs a bit longer.”
You can say what you want about Altman’s tendency to hyperbole but the fact is that OpenAI has traditionally led in model development. Any new competition, especially among US players, in the open model space is welcome. AWS previews AgentCore for enterprise AI agents.
Unfortunately just as OpenAI is preparing to release its first model in six years it was reported Meta, under its expensive new superintelligence laboratory, may abandon its commitment to open-source in favor of a close model.
xAI appears to have already taken this path with its Grok family LLMs. Originally, Elon Musk’s startup planned to release the weights for its last model as soon as a new version came out. While xAI released Grok-1 when Grok-2 was launched, Grok-3 is now out since February and its Hugging Face page looks a bit lonely.
But who would want a model whose hobbies included cosplaying in Mecha Hitler? In this rare case, it may be best to leave this one closed. (r)
