DeepSeek releases V4 open weights under MIT license
2026-08-24
DeepSeek released V4 with full weights on Hugging Face under MIT. Lab-reported numbers put it within five Artificial Analysis index points of the closed frontier, which would be the smallest open/closed gap on record. Too large to run on consumer hardware, but hosted pricing undercuts frontier models by roughly 8x.
Significance 4: First open-weights model within 5 points of frontier on AA index; lab-reported, so capped at 4 until independently reproduced.
Implications · machine-drafted, not owner judgment
Seed example (machine-drafted). If the reported quality holds, near-frontier capability at an 8x price discount collapses the "frontier quality requires frontier prices" assumption that most enterprise AI budgets — and most vendor GTM economics — are built on. It also strengthens the argument that AI capability is commoditizing faster than organizational readiness is improving: the constraint on value shifts further from model access toward deployment competence. What it forecloses: pure model-access resale as a durable business.
- Independent reproduction of the lab-reported scores within 3-4 weeks.
- A frontier lab responding with a price cut rather than a capability release.
- First enterprise deployments citing V4 in place of a closed frontier model.
Sources
- primary DeepSeek V4 release notes retrieved 2026-08-24
- curator Interconnects on the V4 release retrieved 2026-08-24
Models
Affected axes
Hard reasoning · Coding · Cost · Runs well locally on 64GB Apple Silicon