Two years ago, the gap between the best open-weights model and the best proprietary model was roughly the gap between a smart undergraduate and a senior expert. Today, on most tasks that real businesses care about, the gap is closer to that between two senior experts with slightly different specialties. The open ecosystem has matured to the point where 'use a closed model' is no longer the obvious default.
The new heavyweights
Meta's Llama 4 family continues to set the pace for general-purpose open models. The largest variant rivals GPT-5 on many benchmarks, and the smaller variants offer the best quality-per-dollar in the entire industry. Mistral has carved out a strong position with efficient architectures that punch above their parameter count, especially on European languages. Alibaba's Qwen 3 family has become the de facto standard in Asia and is increasingly popular in Western deployments thanks to permissive licensing and excellent multilingual coverage.
Underneath these flagships, a long tail of specialized models — for code, math, biomedicine, legal text — is reaching genuinely production-grade quality. Hugging Face's leaderboards now contain dozens of models that would have been state-of-the-art two years ago.
Why teams are switching
Cost is the most obvious driver. A self-hosted Llama 4 deployment can serve high-volume workloads at a fraction of proprietary API pricing. For startups burning runway on token costs, the math is compelling.
Control is the less obvious but equally important driver. Open weights mean you can fine-tune on proprietary data, audit the model's behavior, run it in air-gapped environments, and avoid the risk of a vendor deprecating the version your product was built on. For regulated industries, that combination is increasingly non-negotiable.
Where open models still lose
Hard reasoning remains the biggest gap. Frontier proprietary models with inference-time reasoning still outperform any open model on competition-level math, complex coding, and long-horizon agent tasks. If your product depends on those capabilities, open models are not yet a drop-in replacement.
Tool use and structured output are another weak spot. Most open models can be coaxed into reliable function calling with careful prompting and fine-tuning, but the out-of-the-box experience is noticeably rougher than with proprietary APIs.
The hosting ecosystem
You no longer need to run your own GPUs to use open models. Together, Fireworks, Groq, Cerebras, and a growing roster of specialized providers offer pay-per-token APIs for the major open models, often at a fraction of proprietary prices and with competitive latency. For most teams, this is the right starting point: get the cost and flexibility benefits of open weights without the operational burden of running your own inference infrastructure.
Fine-tuning is back
For two years, the industry consensus was 'don't fine-tune, just prompt better.' That consensus is shifting. Modern fine-tuning workflows — LoRA, QLoRA, DPO, and their successors — let small teams adapt a base model to a specific domain in a few hours on a single rented GPU. The resulting models are often better and cheaper than prompting a much larger frontier model.
The catch is evaluation. Fine-tuning without a real eval suite is a fast way to ship a worse product. The teams getting the most out of open models are the ones that invest heavily in measurement first.
Practical advice
If you're starting a new project in 2026, the pragmatic stack looks like this: use a frontier proprietary model to prototype and validate the idea, then evaluate whether a hosted open model can serve the production workload at meaningfully lower cost. For most use cases, the answer will be yes.
Keep the proprietary model in your toolbox for the hardest queries, the highest-stakes interactions, and the experimental features where you need the absolute best quality. Use the open model for everything else. That hybrid pattern is becoming the default in production AI systems, and it's a good one.