Open Weights Are the New Open Source

When GPT-3 launched in 2020, the open-source response was GPT-2. The gap was not one generation — it was an entirely different category of capability. The pattern held through 2022 and into 2023: whatever the frontier labs released, open source trailed by a year or more in raw performance, and often by much more on alignment and instruction following.

That story is over. The gap between closed frontier models and the best open-weight releases is now measured in months, not years. On many benchmarks it is gone entirely.

DeepSeek-V3, released in December 2024 under an MIT licence, matched or exceeded GPT-4 on most coding and reasoning evaluations. This was not a small independently-funded lab producing a curiosity — it was a 671B mixture-of-experts model, trained at scale, given away for free. Qwen3 32B and its successor Qwen 3.5 — which extends the context window to 262K tokens and adds refined hybrid thinking across all sizes — hold their own against Claude Sonnet on standard evaluations and run on consumer hardware with 32 GB RAM. Gemma 4 27B ships from Google with open weights and outperforms several proprietary models at twice the parameter count.

The release cadence has also compressed. It used to take six months to a year for the open ecosystem to catch up to a closed model release. Now open-weight responses arrive within weeks. When Anthropic ships a new Claude, the open community is already training the equivalent.

What changed? Compute costs fell. Training techniques matured. The academic and independent research community that produces most of the innovation in this field increasingly publishes with open weights, because that is how you get adoption, citations, and the feedback loops that improve models further. The major frontier labs — Meta, Google, Mistral, Qwen — have concluded that open release is good strategy, not just altruism.

For users, the implication is simple: you no longer have to choose between capability and ownership. Running a model locally does not mean accepting a worse result. In some domains — code completion, document summarisation, structured extraction — local models at the 14B–32B tier are as good as the cloud alternatives that existed a year ago. And they get better every quarter.

The quality moat that justified cloud dependency for serious work has eroded. What remains are the edge cases: tasks that genuinely require the absolute frontier, where an extra two percentage points on a benchmark matters. For the rest of the work most people do with AI, open weights are there.