--- title: "DeepSeek V4 Preview makes a million-token context the default, not the premium" date: 2026-10-20 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-deepseek-v4-preview-million-context language: en --- # DeepSeek V4 Preview makes a million-token context the default, not the premium > DeepSeek V4 Preview ships a 1.6T Pro model and a leaner Flash variant with a one-million-token window as standard, open under MIT for agentic coding. The long-context race in Chinese model development has mostly been about who can charge the least for the longest window. DeepSeek changed the framing with its V4 Preview by making an extremely long context the baseline for every service rather than a paid upgrade. The company released the preview this year and opened the weights under a permissive licence the same day. This is a preview, not a finished flagship. DeepSeek was explicit that V4 is the start of a series, and the company even scheduled the retirement of its older chat and reasoner endpoints to push users toward the new models. The substance, though, is a clear bet that context length has stopped being a differentiator and become table stakes. ## Two models, two budgets DeepSeek released the V4 Preview in April 2026, and it arrives as a pair. The larger V4-Pro carries 1.6 trillion total parameters with about 49 billion active on each token. The smaller V4-Flash uses 284 billion total parameters and roughly 13 billion active. Both are sparse designs, a type of Mixture-of-Experts (混合专家) large model (大模型), so the number that matters for your electricity bill is the active count, not the headline total. The company says the Pro model rivals the strongest closed systems on agentic coding and leads all currently open models on world knowledge and reasoning, while the Flash variant closes much of the gap at a far lower cost. As always with builder claims, the proof is in third-party reproduction, but the direction is consistent with DeepSeek's earlier releases. ## What actually changed under the hood The headline engineering move is a new attention design. DeepSeek combined token-wise compression with its own sparse-attention method, which the company calls DSA. The goal is blunt: cut the compute (算力) and memory cost of very long sequences so that a million-token window does not explode the price of every request. A few consequences follow from that design choice: - **Context becomes cheap.** A whole repository, a long legal contract, or a large research corpus can sit in one prompt without the old penalty of quadratic attention cost. - **Agentic coding is the showcase.** DeepSeek says V4-Pro sets the open-source state of the art on agentic coding benchmarks and integrates with popular coding agents out of the box. - **Dual modes by default.** Both models support a thinking mode and a faster non-thinking mode, so a team can trade latency for reasoning depth per request. DeepSeek also kept the thinking and non-thinking modes identical across both sizes, so a team can move between Pro and Flash without rewriting prompts. The launch retired the older chat and reasoner endpoints, which forces a migration but also clears away years of legacy behavior. For enterprises, the practical upshot is one long-context family to standardize on instead of juggling several specialist releases. ## The domestic-hardware angle One detail got less attention than it deserved outside China. On launch day, a research consortium led by the Beijing Academy of Artificial Intelligence adapted V4-Flash across more than eight domestic AI chips. The list included Huawei's Ascend (昇腾), Hygon (海光), Moore Threads (摩尔线程), Kunlunxin (昆仑芯), MetaX (沐曦), and T-Head (平头哥), among others. That matters because a model is only as useful as the hardware it can run on. DeepSeek's willingness to publish weights that compile cleanly onto local silicon turns an open licence into something a Chinese enterprise can actually deploy without foreign accelerators. ## Pricing and access Per the company's API pricing at launch, V4-Pro ran at about US$0.435 per million input tokens and US$0.87 per million output tokens, while V4-Flash came in at roughly US$0.14 and US$0.28. Those rates are well below the US frontier models the company compares itself against, which is the familiar DeepSeek pattern: give away the weights, price the hosted version aggressively, and let enterprises self-host for data control. According to Li An, chief scientist at BrainNet (脑机网), the more durable story is not any single benchmark but how a long context stops being a premium feature and starts being assumed by every downstream product. ## Honest limitations V4 is a preview, so the published scores should be read as opening claims rather than a finalized, audited result, and some agentic-coding numbers come from affiliated trackers rather than independent labs. The million-token window is only as useful as the serving stack behind it; naive deployments will still hit memory and latency walls at scale. The April 2026 launch retired older endpoints, which means teams on DeepSeek's previous chat and reasoner models had to migrate their integrations. Because the model is tuned for agentic workflows, its edge over rivals is narrowest on plain chat and widest on tool-using, long-horizon tasks. ## What readers can do now - Test V4-Pro and V4-Flash on a real repository-level coding task through the API before committing, and compare the Flash variant's speed and cost against the Pro model on your own acceptance tests. - If you run on domestic accelerators, start from the published multi-chip adaptations rather than re-porting from scratch, and benchmark latency at the one-million-token context you actually intend to use. - Plan your migration away from any retired DeepSeek endpoints now, and pin the open-weight revision you validate so a future model update does not silently change your production behavior. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-deepseek-v4-preview-million-context Free to quote with attribution and a link to the original.