DeepSeek V4 Preview lands as an architectural reset — V4-Pro (1.6T total / 49B active) targets long-context reasoning, V4-Flash (284B / 13B) targets cost
DeepSeek released V4 Preview on April 24, splitting into two SKUs: V4-Pro at 1.6T total / 49B active parameters for long-context reasoning and agentic workflows, and V4-Flash at 284B / 13B active for cost-efficient deployment. The release is an architectural reset rather than an incremental update, and it lands in a quarter where Qwen 3.6, Llama 4, Gemma 4, and Mistral Medium 3.5 are also pushing the open-weight frontier.
The split-SKU strategy mirrors the closed-lab playbook — top-of-line for capability, smaller variant for cost — but applied to open weights. V4-Pro's 49B active parameter count with 1.6T total is the kind of mixture-of-experts scale that previously required proprietary tooling. Releasing it as open weights closes another gap with closed labs and tightens the calibration question Dario Amodei has been raising publicly about Chinese-lab efficiency claims.
The competitive reading: open-weight models are now keeping pace with closed-weight ones on capability and shipping faster on architecture. The open-weight community has gone from "trailing" in 2023 to "adjacent" in 2025 to "leading on benchmarks" in pockets of 2026 (Qwen 3.6 Max-Preview holds #1 on six major coding and agent benchmarks including SWE-bench Pro and Terminal-Bench 2.0). The closed labs no longer have the capability moat they had two years ago.
LLM-Stats — AI Updates Today May 2026 Latest AI Model Releases → · Codersera — Best Open-Source LLM in May 2026 → · Digital Applied — Open-Weight Models H1 2026: DeepSeek, Qwen, Llama Recap →