A lot of people throw around the word “free” when they talk about open-weight LLMs. Technically, they are not wrong. You can download the weights, inspect them, fine-tune them, and often use them commercially. But that is only half the story. The part that matters in practice is this: a model can be free to access and still be expensive, awkward, or outright unrealistic to run on normal hardware.
That is the gap this article is really about.
The latest wave of open-weight models makes the contradiction even more obvious. Qwen has released open-weight families that scale from smaller dense and MoE variants to huge flagship models, including Qwen3.5 with an open-weight first model in the series, Qwen3.6-35B-A3B, Qwen3.6-27B, and Qwen3-Max, which the company says has over 1 trillion parameters. DeepSeek has also kept pushing the same pattern, with DeepSeek-R1 released under MIT, DeepSeek-V3-0324 also under MIT, DeepSeek-V3.1 and DeepSeek-V3.2 as open-source releases, and DeepSeek-V4 Preview advertising 1.6T total parameters with 49B active in one variant and 284B total with 13B active in another.
GLM follows the same playbook. GLM-4.5 is an MoE model with 355B total parameters and 32B active parameters per forward pass, while GLM-4.5-Air uses 106B total and 12B active. Z.ai also says GLM-5.1 is open source under the MIT License. Kimi’s newer open-source coding models keep the scale high too. Kimi K2.7 Code is listed as a 1T total, 32B active MoE model with 256K context, and Kimi K2.6 is also positioned as an open-source model for coding and agent work. MiniMax is not any different. MiniMax-M2.1 is described as a 230B total, 10B active MoE model, and MiniMax-M3 is a frontier coding and agentic model with 1M context.
That is where the “free” story starts to crack.
A model with hundreds of billions of parameters is not just a bigger download. It is a different deployment class. It changes memory pressure, throughput, latency, batching strategy, quantization tradeoffs, and the number of users you can serve before the economics stop making sense. In other words, open weights remove the licensing barrier, but they do not remove physics. If the model wants 32B active parameters or more, the average developer is no longer asking, “Can I use this?” They are asking, “Can I afford to keep this alive?” The answer is often no, at least not on a normal consumer GPU setup. That is not a licensing issue. It is a systems issue.
This is why the word “efficient” needs to be defined more carefully.
Efficient does not mean “smallest possible.” It means the best ratio of capability to deployment cost. A genuinely efficient model is one that gives you enough reasoning, coding, tool use, or multilingual quality without demanding a data center just to answer a few prompts. That is the real benchmark. Not raw parameter count. Not marketing. Not whether the weights are technically available on Hugging Face.
The industry keeps rewarding scale because scale still works. But scale has a tax. The bigger the model gets, the more it becomes a strategic asset for labs and a logistical headache for everyone else. You can see why many of these releases lean on MoE. A model like GLM-4.5 or Qwen3.5 can advertise very large total capacity while activating only part of it per token, which improves serving efficiency on paper. The catch is that “more efficient than a dense model of the same total size” is not the same thing as “easy to run locally.”
So what should builders actually take from this?
Use open weight models, absolutely. They are one of the best things to happen to the ecosystem in years. But choose them with deployment reality in mind. If you need local inference, look for smaller dense or smaller MoE variants first. If you need a coding agent, compare active parameters, context length, and real-world latency, not just benchmark headlines. If you are shipping a product, ask whether the model fits your infra budget before you ask whether it tops a leaderboard.
That is the practical lesson here.
Open weight gave us access. It did not give us free compute. And that distinction matters because it changes how you build, how you deploy, and which models are actually usable outside a lab. The best open-weight model is not the largest one on the announcement page. It is the one that delivers enough quality without turning your hardware into the bottleneck.
Top comments (0)