The open-source LLM landscape in May 2026 looks very different from twelve months ago. The model quality gap between the best frontier closed models and the strongest open-weight models has narrowed but hasn’t closed. The deployment economics of open models have improved enough that enterprise production usage has gone from a small cohort of experimenters to a meaningful fraction of total inference spend.

The models actually running in production right now: Llama variants remain the most-deployed open base model, particularly the post-training derivatives that have gone through serious instruction-tuning and reinforcement learning steps. Mistral’s larger models hold real ground in European enterprise deployments. Qwen has built a meaningful production presence, especially in deployments where multilingual performance matters. The smaller specialist models — Phi family, smaller Gemma derivatives — are having a moment in edge and on-device applications.

What’s changed since late 2025 is the maturity of the post-training tooling around these models. Fine-tuning pipelines, distillation workflows, and continuous evaluation frameworks have all gotten meaningfully better. The teams running open models in production now generally have a continuous evaluation harness running against their production query distribution, with synthetic data generation and targeted fine-tuning loops feeding back into model updates every few weeks.

The economic case for open models is strongest in two specific places. First, high-volume inference at predictable load: when you’re running enough tokens per day that the API costs of frontier closed models become a meaningful line item, the operating cost of self-hosted open models is dramatically lower. Second, deployments with hard data residency or sovereignty requirements, particularly in regulated Australian sectors. The open-model deployment story for Australian healthcare, financial services, and government has become far more pragmatic than it was even a year ago.

The economic case is weakest for low-volume usage and for use cases that need the absolute frontier of capability. The cost of running an open-model serving infrastructure — H100s or equivalent, ops staffing, evaluation tooling — only makes sense at meaningful scale. For startups and smaller teams, the rational play is still API-served frontier models for reasoning-heavy work, and possibly a smaller open model for specific high-volume sub-workloads.

Quantisation is where the practical engineering action is. The gap between full-precision and quantised open-model performance has narrowed enough that 4-bit and 8-bit quantised inference now produces production-acceptable quality on most enterprise workloads. The serving cost difference is significant. Most production deployments running open models in 2026 are running them at 4-bit or 8-bit quantisation by default, with full precision reserved for specific evaluation or accuracy-sensitive paths.

For Australian enterprises planning open-model deployments, the practical questions are: do we have the operational maturity to run model serving infrastructure, do we have a credible evaluation harness, and is our workload at the scale where the economics actually flip. Most enterprises that ask those questions honestly conclude that they should run open models for some workloads and continue using API-served frontier models for others. The pure-open or pure-closed positions are mostly ideological.

The trajectory of the next twelve months looks like more of the same. Open models will continue to close the capability gap on specific tasks. Closed models will continue to lead on the hardest reasoning and agentic workloads. The hybrid pattern — open models for some workloads, closed models for others, intelligent routing between them — is the production architecture that’s actually winning in 2026.