4 open-weight models
Open-weight models
Models whose weights you can download and run on your own hardware. The capability gap to closed frontier models narrowed sharply during 2026, and the current generation is trained for agent use rather than just chat.
Pricing checked against provider documentation on . How we verify
Open weights ≠ open source
Downloadable weights do not imply a permissive licence, and they do not imply released training data or code. Several widely used "open" models restrict commercial scale or competitive use. Read the licence before you build on one — the prices below are for hosted API access, which is a separate question from whether you may self-host.
4 of 4 models
| Model | Provider | Input / 1M | Output / 1M | Context | Best for | Status |
|---|---|---|---|---|---|---|
| DeepSeek V4-Flash reasoning open | DeepSeek | $0.22 | $0.66 | 1M | The cheapest capable open-weight option | stable |
| DeepSeek V4-Pro reasoning open | DeepSeek | $0.66 | $1.98 | 1M | Frontier-adjacent reasoning at open-weight prices | stable |
| GLM-5.2 reasoning open | Z.ai | $1.40 | $4.40 | 1M | MIT-licensed coding agents | stable |
| Kimi K3 reasoning open | Moonshot AI | $3.00 | $15.00 | 1.05M | Natively multimodal long-horizon agents | stable |
No models match those filters.
Providers
DeepSeek, Moonshot AI, Z.ai
Licences in play
MIT, Kimi K3 License (commercial terms above $20M revenue)
Cheapest hosted access
Utilisation decides the maths
A GPU costs the same idle as busy. Self-hosting wins on steady high throughput and loses badly on spiky traffic — model your duty cycle, not your peak.
Data residency is often the real driver
Teams frequently choose open weights for control rather than cost: on-premise deployment, no third-party retention, and auditability that hosted APIs cannot offer.
Hosted open models are a middle path
Inference providers serve open weights per-token, giving you portability between vendors without running infrastructure. Cheaper than closed frontier models, no GPU commitment.
Frequently asked
Does 'open weights' mean the model is free?
No. It means the weights are downloadable so you can run the model yourself. You still pay for GPU time when self-hosting, or per-token fees when calling it through a hosted provider. 'Open weights' also does not mean open source — training data and code are usually not released.
When does self-hosting actually save money?
At sustained high volume. You pay for the GPU whether or not requests are arriving, so utilisation is everything. Spiky or low traffic is almost always cheaper on a hosted API. Run the arithmetic including engineering time before committing.
How far behind are open models now?
Much closer than in 2024. Open releases narrowed the gap considerably during 2026, and crucially they are now trained for agent use — tool calling, structured output, and long-context reasoning designed in rather than bolted on. The remaining gap is largest on the hardest reasoning tasks.
What licence should I check for?
Read the actual licence rather than assuming. Several popular 'open' models carry restrictions on commercial scale, competitive use, or attribution. Apache 2.0 and MIT are permissive; bespoke community licences frequently are not.