Open-weight models
DeepSeek ships V4: 1.6T open-weight model, $1.74 a million tokens
The brief: two MIT-licensed MoE models, 1M context, and the first DeepSeek architecture built for Huawei chips.
The answer
DeepSeek released V4-Pro (1.6T params) and V4-Flash (284B) on 24 April 2026 under MIT.
What happened
On 24 April 2026 DeepSeek released two Mixture-of-Experts models under MIT: V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B / 13B active). Both support a 1 million-token context window — up from 128K in V3.2 — and a 384K max output. Weights are live on Hugging Face; the API went live the same day. The legacy deepseek-chat and deepseek-reasoner model IDs will be retired 24 July 2026, mapping to V4-Flash in the interim. A new sparse-attention architecture (CSA + HCA) cuts V4-Pro's inference FLOPs to 27% of V3.2 and KV-cache to 10%.
The numbers: benchmarks and price
Artificial Analysis placed V4-Pro #2 on the open-weight reasoning index (behind Kimi K2.6) and leading open models on agentic coding. Standard API pricing: V4-Pro $1.74/M input / $3.48/M output (launched on a 75% promo at $0.435/$0.87); V4-Flash $0.14/M input / $0.28/M output. CFR put the overall U.S. AI lead at roughly seven months — but noted the competitive contest is adoption scale, not just leaderboard rank.
| V4-Pro | V4-Flash | V3.2 (prev.) | |
|---|---|---|---|
| Params (total / active) | 1.6T / 49B | 284B / 13B | 671B / 37B |
| Context | 1M tokens | 1M tokens | 128K tokens |
| API input price (list) | $1.74/M | $0.14/M | $0.27/M |
| Inference FLOPs | 27% of V3.2 | ~10% of V3.2 | baseline |
CFR fellow Michael Horowitz observed that 'second-best models carry enormous competitive value when they are cheap and open, which makes them easy to widely diffuse' — reframing the contest from benchmark supremacy to global deployment reach.
What to watch. Whether DeepSeek publishes a technical paper (the efficiency claims are plausible but currently self-reported). Whether inference on Huawei Ascend proves cost-competitive with Nvidia at scale. Whether V4's MIT licence accelerates global adoption outside the U.S. — the strategic variable the leaderboard doesn't capture.
Frequently asked questions
What did DeepSeek release on 24 April 2026?
How much does DeepSeek V4 cost via API?
Is DeepSeek V4 better than GPT-5 or Claude?
Sources
- DeepSeek V4 Preview Release — DeepSeek, 24 April 2026
- Three reasons why DeepSeek's new model matters — MIT Technology Review, 24 April 2026
- DeepSeek is back among the leading open-weights models with V4 Pro and V4 Flash — Artificial Analysis, 27 April 2026
- DeepSeek V4 signals a new phase in the US–China AI rivalry — Council on Foreign Relations, 29 April 2026
- DeepSeek previews new AI model that 'closes the gap' with frontier models — TechCrunch, 24 April 2026