4 articles
Browse guides
DeepSeek released V4.1 Flash on September 10, 2026 and is routing its own V4 Pro flagship to it four days later at Flash prices. A 552B backbone that activates 8B parameters to read and 16B to write, native vision, an 890-byte-per-token KV cache, and a scaffold table showing the same model swing 8.7 points on one benchmark depending on the agent harness.
Abliteration.ai sells token-priced access to open-weight models whose refusals have been suppressed in the weights. What TechCrunch’s testing showed, why researchers are split on how much it matters, why open weights cannot be recalled, and the practical rules for companies and everyone else.
DeepSeek’s risk depends on what you send and how it’s deployed: hosted convenience versus self-hosted control, with personal and regulated data kept out either way.
DeepSeek V4-Flash-0731 is an official, open-weight, MIT-licensed MoE model with a 1M-token context, tuned for agents and coding, at about $0.14 per million input tokens. Specs, benchmarks, and why open weights plus low prices pressure the closed labs.