Ahmad Osman predicts that GLM 5.2-level intelligence will run on one RTX 5090 with 32 GB of VRAM within roughly 18 months.
2
Model research is producing more capability per parameter, so hardware requirements are falling even as local models gain reasoning, context, and agentic abilities.
3
Owning local hardware gives individuals and businesses control over their models, costs, and future model availability.
Summary
Ahmad Osman argues that frontier-level AI is moving onto personal hardware because model efficiency is improving faster than model size is growing. He uses the idea of impact per parameter to compare older systems with newer models that deliver similar or better capabilities using fewer parameters and smaller hardware footprints. He points to the progression from Llama 2 and Mistral through Qwen, DeepSeek R1, and newer tool-calling models as evidence that open-weight systems are narrowing the gap with cloud frontier models. Osman also makes a case for sovereign AI. Individuals and companies can own the compute stack, tune models for their own work, and avoid depending entirely on cloud providers. His final argument is economic: hardware bought today may become more useful as increasingly efficient models arrive, so he is keeping his RTX 3090s rather than selling them.
A single RTX 5090 could run GLM 5.2-level intelligence within roughly 18 months
Osman predicts that a single RTX 5090 with 32 GB of VRAM will run the equivalent of GLM 5.2-class intelligence within roughly 18 months, which he places around late 2027. He calls this a conservative estimate and says it could happen sooner. The prediction depends on the shrinking gap between frontier systems and open models, plus continuing gains in model efficiency. He does not claim that the gap will disappear. His point is that the hardware needed for a given level of capability is falling quickly.
Impact per parameter is a better way to track local model progress
Osman asks readers to compare what a model can do with the hardware footprint it needs. He describes this as impact per parameter. He once ran Llama 2 on an RTX 3090, while a 27-billion-parameter Qwen 3.5 model now provides better capability than a Llama model with more than 400 billion parameters. He uses this comparison to show that model size alone does not explain progress. Similar capabilities are moving into smaller systems, and local models are gaining longer context and more useful behavior at the same time.
Architecture improvements compound over successive model generations
The move toward smaller local systems is not random, according to Osman. Research gains, training improvements, and architecture changes compound across releases. He connects this pattern to the Densing Law, which he attributes to Nature Machine Intelligence. His description is that the number of parameters needed for a given level of intelligence falls by about 50 percent every three and a half months, although dense and activated parameters are different measures. The practical result is more capability from the models people can run on hardware they own.
Frontier-class open models are already running on desktop-scale systems
Osman says GLM 5.2 has 744 billion total parameters with 40 billion activated and supports contexts of up to one million tokens. In MVFP4, he says it can run on a DGX Station or on a server with eight RTX Pro 6000 GPUs. A DGX Station can sit under a desk while running this kind of model. He also cites Neatron 3 Ultra as evidence that more efficient training can happen on available hardware, which could lower the cost of fine-tuning and building smaller, specialized models.
Sovereign AI gives users control over models and compute
Osman argues that consumers, small businesses, and enterprises should consider sovereign AI, meaning they control the models and hardware they use. Local ownership can allow later optimization for specific use cases and reduce ongoing costs. He also says open AI needs enterprises to own their hardware rather than only paying cloud providers. That ownership would create demand for open-weight models and support an ecosystem in which providers continue releasing models and developing licenses that let open source continue.
Open-weight models have moved from basic local use to reasoning and tool calling
Osman traces progress from Mistral 7B and Llama 3 through Qwen 2.5 and DeepSeek R1. He says DeepSeek R1 made reasoning practical at home, although its roughly 671-billion-parameter mixture-of-experts design still required a substantial server. He then points to GPT OSS 12B as one of the first open models that could successfully do tool calling. This reduced hardware footprint brought more agentic behavior to home systems. Later comparisons between larger mixture-of-experts models and dense 27-billion-parameter models show how quickly this progression is happening.
The value of a GPU may rise as more capable models become smaller
Osman questions whether a hardware purchase can become more useful as model efficiency improves. He is keeping his RTX 3090s because he wants to see what they can run in one or two years, rather than selling them for their current value. He notes that RTX 3090 cards still sell above their original MSRP and remain useful for many workloads. He extends the same question to the DGX Station, asking what the actively developed Blackwell architecture will be able to run in six, 12, or 18 months.
"Why wouldn't you want to be in control of the models that you run? Why wouldn't you want to make sure that nothing gets taken away from you?"07:59
Who should watch
You are deciding whether to run models locally or continue paying for cloud inference.
Your team needs private, controllable models and is considering buying or keeping GPU hardware.
You follow open-weight models and want a concrete account of how reasoning, tool calling, and model efficiency have changed.
