Event Summary
On July 31, 2026, DeepSeek promoted DeepSeek-V4-Flash-0731 from preview to an official public beta API, retaining its 284B parameter MoE architecture while improving post-trained agent capabilities and cost efficiency.
Context & Narrative
DeepSeek released an updated official API public beta checkpoint named DeepSeek-V4-Flash-0731. The model retains its 284 billion total parameter Mixture-of-Experts architecture (13 billion active parameters per token) while integrating a fresh round of fine-tuning that improves tool-use, terminal coding, and long-horizon reasoning. Available via DeepSeek's official API at $0.14 per million input tokens, the release also added native support for OpenAI's Responses API format and local Codex execution.
Key Findings
-
Fact Grade B
DeepSeek transitioned DeepSeek-V4-Flash-0731 from preview to an official public beta API on July 31, 2026, featuring 284B total parameters and 13B active parameters at $0.14 per M input tokens.
-
Impact Grade C
Provides an open-weight lightweight MoE checkpoint achieving high intelligence scores on low inference cost constraints.
Sources [3] -
Limitation Grade B
The update was limited to the V4-Flash model checkpoint; V4-Pro and web application models remained unchanged until mid-August.
Sources [1]
Impact Assessment
-
Access Democratization +1 · Medium-term
Provides high-throughput open-weight MoE model API at ultra-low inference costs.
Affected Groups: developers, AI developers, open-source community
Consensus & Sources
-
1
Reference Evidence Citation logged Live source
-
2
Reference Evidence Citation logged Live source
-
3
News Report Citation logged Live source