Review update · 17 October 2026. Corrected model sizes, modalities and memory assumptions; removed unsupported parity claims. This remains a historical 2025 snapshot, with separately dated token guidance.
This article is a corrected snapshot of the open-weight landscape discussed in March 2025, not a ranking of the best models available today. The useful questions survive longer than the leaderboard: what can you run, what are you permitted to do with it, and does it perform on your actual work?
Correcting the old categories
The earlier tier list placed Llama 3.2 11B and 90B together in a 7B–34B category. That was wrong: 90B is outside that range, and these are vision models accepting text and images, not simply two text-only sizes. Meta’s model card documents their modalities and sizes.
DeepSeek-V3 also needs more than a “70B+” label. Its technical report describes a mixture-of-experts model with 671 billion total parameters and 37 billion activated per token. Active parameters help explain computation; they are not the total weights you must accommodate when deploying the full model.
There is no sound basis here for the old claim that open models have closed most quality gaps across most production tasks. A benchmark result applies to a particular model, test and setup. It does not establish parity on private documents, long tool workflows or a different language.
Weights are only part of the memory budget
As arithmetic, 70 billion weights at four bits each occupy about 35 billion bytes before quantisation metadata and other overhead. That is not a GPU sizing recommendation. The KV cache, context length, batch size, activations and runtime also need memory; offloading changes latency and throughput. Measure the exact deployment rather than promise that a model fits from its parameter count alone.
Quantisation format and method are separate choices. GGUF is a file format used for model distribution, whereas AWQ and GPTQ describe quantisation methods. Reduced storage does not guarantee unchanged quality; test the quantised artifact you intend to serve.
Open weights do not settle the licence question
Accessible weights do not automatically mean unrestricted use. Review the licence for the exact artifact and version, including attribution, redistribution, acceptable-use and commercial conditions where applicable. The linked Meta and Qwen cards identify different licence families. Treat this as a deployment check, not legal clearance from this article.
A more useful evaluation
- Build a representative set of tasks with expected outcomes and unacceptable failures.
- Compare the exact quantisation, prompt, context length and serving setup you would deploy.
- Measure latency under load and memory at realistic concurrency, alongside quality and cost.
- Check access controls, logging, data residency, patching and the time needed to operate it.
- Retest after changing the model or runtime. A valid JSON response can still contain incorrect facts.
Self-hosting is an operational choice. It can provide control over where processing occurs, but it does not automatically make a system private, cheaper or safer. Those outcomes depend on the surrounding system and the workload.
Related reading: LLM Inference Costs: The Race to the Bottom and What It Actually Means.