For AI on your own infrastructure what counts is the gap to the frontier - and whether the model can be swapped.
What it is about
The best open models lag the leading closed ones by about four months on average, by six to seven for cyber capabilities. Models for a single graphics card are six to twelve months behind.
Organisations that may only run AI in their own data centre work with open models. Epoch AI measures the gap continuously with its own capability indices - an independent research institute, not a model vendor. These are benchmark analyses, not experiments with users.
What the benchmarks show
- Since January 2026 the best open models have been four months behind the leading closed ones on average (Epoch Capabilities Index, gap of 8 points). In October 2025 the gap was about three.
- For cyber capabilities the gap is 6-7 months (Epoch’s cyber index, October 2026).
- Models that run on a single consumer graphics card reach frontier level with a delay of 6-12 months (2025, benchmarks GPQA Diamond, MMLU-Pro, AA Intelligence and LMArena).
The best open models are often too large for typical in-house hardware, and small models are partly tuned for benchmarks. In practice the gap can therefore be larger.
What we make of it
For a realistic case - a server with eight GPUs - we analysed the gap ourselves: 3-5 months for coding, 8-10 months without models of Chinese origin. Details in On-premises AI: three to ten months behind the cloud. The consequence: plan the architecture so that the model stays replaceable.
The studies in detail
Selection: only studies with a disclosed method - experiments, official statistics, legal texts, independent institutes. Vendor surveys about their own product and consultancy forecasts are not included here.
- 4 Monate
Epoch AI · 2026 · Benchmarkindependent
Since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months (Epoch Capabilities Index, gap of 8 points).
Limits: Measures the best open-weight models, which are often too large for typical on-prem hardware. Epoch's Oct 2025 estimate was about 3 months.
Open models lag state-of-the-art closed models by 4 months ↗We use it in: On-premises AI: three to ten months behind the cloud
- 6-7 Monate
Epoch AI · 2026 · Benchmarkindependent
In cyber capabilities, open-weight models trail closed models by 6-7 months (Epoch AI, Cyber ECI, Oct 2026).
Limits: Cyber-specific index, so it does not carry over to general tasks.
Open-weight models trail closed models in cyber capabilities by 6-7 months ↗We use it in: On-premises AI: three to ten months behind the cloud · TISAX meets GenAI
- 6-12 Monate
Epoch AI · 2025 · Benchmarkindependent
Models that run on a single consumer GPU (e.g. RTX 5090) match frontier-model performance with a lag of 6 to 12 months (Epoch AI, 2025).
Base: Benchmarks GPQA Diamond, MMLU-Pro, AA Intelligence, LMArena
Limits: Epoch notes the real-world lag may be longer, because small models are often tuned for benchmarks.
Frontier AI performance becomes accessible on consumer hardware within a year ↗We use it in: On-premises AI: three to ten months behind the cloud
