What are open-weight models, and how far behind the frontier are they?

An open-weight model is one whose trained parameters you can download and run yourself under a published licence, which is not the same thing as open source, because the training data and the training pipeline almost always stay behind the door. On the hardest public benchmark we can price properly, the best open-weight configuration currently scores about 58 percent of what the leading closed model scores, and costs about 44 percent as much per point of that score.

Those two numbers are most of the argument.

Everything below is where they come from, what the licences do to them, and when they are likely to move.

Open weights is a licence on a file, and open source is a much longer list

The Open Source Initiative's definition asks for three things: Data Information, Code and Parameters. It wants "sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system", plus "the complete source code used to train and run the system", and it says plainly that "'Open Source models' and 'Open Source weights' must include the data information and code used to derive those parameters".

Almost nothing marketed as open weights clears that bar. DeepSeek's own card for V4.1-Flash is unusually generous by industry standards — it describes a 552B-backbone mixture of experts activating "only 8B parameters per token during prefill and 16B during decode", trained "from scratch on a multimodal corpus comprising 45T tokens" — and still never names the corpus. You get the recipe's proportions. You do not get the ingredients.

The exception is worth knowing about. OpenBMB shipped MiniCPM5-2B under Apache 2.0 and released the training data alongside it, as "UltraX, a high-quality web pre-training dataset; UltraData-Code … UltraData-SFT-Agent-2609 … and UltraData-RL-2609". That is a 2.5-billion-parameter model clearing a bar none of the frontier-scale releases in this article clear.

Four models, four different licences

Check the file before you check the benchmark.

Model	Licence	The catch
DeepSeek V4.1-Flash	MIT	None
MiniCPM5-2B	Apache 2.0	None, and the training datasets ship too
Kimi K3	Kimi K3 License	Separate agreement with Moonshot if you run model-as-a-service above $20m revenue
GLM-5.3	GLM-5.3 License	Z.AI security review if you run model-as-a-service above $10bn revenue

Kimi's threshold is the one that bites. Its licence defines model-as-a-service as "giving a third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data", excludes "end-user products with model capabilities solely embedded within specific features", and then requires a separate agreement once revenue passes twenty million dollars over any twelve months. Zhipu's clause is the same shape with the trigger set five hundred times higher, at ten billion, and a security review instead of a negotiation. Both reach derivative works.

Neither is MIT, whatever the download button suggests. If your company might plausibly resell inference, read clause two before you read the benchmark table.

Running one is a memory problem before it is a model problem

Two paths. Serve it on GPUs with vLLM or SGLang, which is what the model cards assume — Z.ai publishes vLLM recipes and an SGLang cookbook for GLM-5.3, and OpenBMB ships MiniCPM5-2B in "GGUF for llama.cpp / Ollama / LM Studio". Or rent it by the token from a host, which is probably how most people actually use open weights. Ollama's cloud lists glm-5.3-flash at $0.15 per million input tokens and $0.50 output. GPT-6 Astra is $10 and $50.

That ratio is why open weights keep turning up in the parts of a stack nobody puts in a slide.

The third path is quantisation, and it is where the interesting engineering is. vLLM's documentation states the trade in one line: "Quantization trades off model precision for smaller memory footprint, allowing large models to be run on a wider range of devices." The naive version quantises everything and loses the model. Salvatore Sanfilippo, who has been doing this in public, published the working version on 27 April 2026: "drastically quantize only routed experts, in/gate IQ2_XXS, out Q2_K. Then: leave all the rest Q8, to preserve quality in shared experts, projection, routing."

By 28 August he had GLM 5.3 Flash running at Q2 and Q4 on "single MacBook 128GB or DGX Spark inference, two MacBook RDMA 128GB each Q4 inference in tensor parallel fashion at 37 t/s single generation". Thirty-seven tokens a second across two laptops.

That is roughly the state of the local frontier, and we wrote it up in GLM 5.3 reaches Ollama cloud-only, while antirez runs it at Q2 on two Macs.

The gap, measured, with the cost divided out

Vercel's DeepsecBench scores models on finding real vulnerabilities in application code and publishes the dollar cost of every run, which makes it the rare leaderboard you can do arithmetic on. The cost-per-point column is ours.

Model (best configuration)	Score	Run cost	Cost per point
GPT-6 Astra, xhigh	37.79	$63.70	$1.69
GPT-5.6 Sol, xhigh	35.44	$55.98	$1.58
Claude Opus 5, max	32.44	$127.93	$3.94
GLM-5.3, high	21.91	$16.22	$0.74
Kimi K3, xhigh	17.49	$12.38	$0.71
DeepSeek V4 Flash, xhigh	16.47	$5.94	$0.36

GLM-5.3 is 58 percent of Astra's score at 44 percent of its cost per point. DeepSeek V4 Flash is 44 percent of the score at 21 percent of the cost per point, which makes it the cheapest capability on the board by a distance nobody advertises. Vercel's own summary is fair about the direction: "Frontier models from OpenAI and Anthropic still score highest, but open-weight models and more efficient reasoning options are closing that gap, making comprehensive scanning far more cost-efficient."

Is the gap closing or widening? Both, on different axes, which is the argument we walked through in Open weights sit at 58 percent of the frontier and 44 percent of the price. Score gaps seem to have widened since March. Money gaps have narrowed.

Companies build on them because the second number is the one that ships

Harvey, valued at $15.5bn on 9 September 2026, built its first in-house model on Moonshot's weights. In its own words, "Harvey Tenet is a Kimi K3 base that we post-trained together with Fireworks research for long-horizon legal work", and the post-trained model "successfully completes almost twice as many held out tasks on LAB and 20% more on LAB contracts than base Kimi K3".

The reason it gave is the one that matters, and it is not ideology. "Open-weight models have cheaper per token prices. But cost is a function of both token prices and tokens used." Post-training let Harvey shape the second term, which you cannot do through somebody else's API at any price. That trade, and the licence question underneath it, is Harvey raises $550m at $15.5bn, and its own model runs on Chinese weights.

The counter-argument is real. A 2.8-trillion-parameter model with 104 billion active is not something you run on your own hardware, so for most teams open weights means renting the same tokens from a different vendor, with the freedom to move as the actual benefit. Fair. Portability is probably worth less than people claim and rather more than nothing, and as far as we can tell it is the only reason the price column above stays honest.

What would change the answer

Four things, roughly in order of likelihood. An open-weight model breaking into the top three of a cost-published leaderboard, which would end the 58 percent framing. A licence change at Moonshot or Z.ai that moves a revenue threshold downwards, which would end the free-in-practice framing. A Western lab shipping frontier weights, which nobody has. And a single consumer machine holding a 300-billion-parameter model at four bits without a second machine bolted to it, which is closer than it was in April and still not here.

More in our open-weights coverage.