DeepSeek-V4 officially launches, adding Pro and Flash versions to its API
Over the past two years, DeepSeek has maintained a rapid pace of model releases: V3 launched in late 2024, R1 followed in early 2025, and a series of upgrades—including V3.1 and V3.2—came afterward. In February 2026, some users got an early look through an app update at a gradual rollout that expanded the context window from 128K to 1M, with the knowledge base updated through May 2025. On April 24, the official changelog confirmed the formal launch of DeepSeek-V4.
The official changelog shows that the API now supports two new versions, V4-Pro and V4-Flash, and is compatible with both the OpenAI ChatCompletions API and the Anthropic API. The endpoint remains unchanged; developers only need to set the model parameter to deepseek-v4-pro or deepseek-v4-flash. The old model names deepseek-chat and deepseek-reasoner will be discontinued on July 24, 2026. During the transition period, they will respectively point to V4-Flash’s non-thinking mode and thinking mode. A June report from a third-party cloud platform also said that the production version of V4 introduced a 1M-token context window and a new sparse attention architecture.
A closer look at this update suggests a clear emphasis on efficiency rather than parameter-count showmanship. A previous Nomura Securities report said that two design features introduced in V4—mHC manifold-constrained hyper-connections and the Engram conditional memory module—address two major pain points: instability in deep-network training and high memory usage. The former makes information flow between layers more stable, while the latter places static knowledge in inexpensive memory, freeing up costly HBM. The firm believes V4’s core value lies in using architectural innovations to lower training and inference costs and advance the commercialization of AI applications—not in upending the existing AI value chain. Nor does it expect V4 to recreate the panic over computing demand seen when V3 launched. Notably, the report also mentioned that DeepSeek models accounted for more than half of calls to open-source models on OpenRouter in late 2024, but their share had fallen significantly by the second half of 2025. V4 faces a much more complex competitive environment than it did a year ago.
The impact on crypto markets can only be assessed through logical inference for now; the available evidence contains no data on how related sectors actually responded after V4’s release. There are three possible transmission channels. First, if inference costs fall substantially, the demand narrative for decentralized computing projects could be reassessed, echoing the institutional view that the release will not trigger panic over computing demand. Second, low-cost long-context capabilities could benefit AI agent development: multi-step tasks consume more tokens, potentially providing a catalyst for the agent narrative. Third, as open-source models continue to approach the performance of closed-source models, expectations for local deployment and decentralized inference may strengthen. These are all projections, not established facts.
Editor's note: The clearest signal in this update is the deliberate effort to keep migration costs low. Compatibility with both APIs, combined with a three-month transition period for the old model names, shows that the official priority is the developer ecosystem. For the crypto industry, the ultimate beneficiaries of improved model efficiency are more likely to be application-layer projects that generate real usage and on-chain settlement demand, rather than assets linked to AI only by association. Two sets of data are worth tracking: migration activity after the old model names are discontinued on July 24, and actual changes in application-level usage.
#DeepSeek发布更新版AI模型 #BTC #ETH #BNB
Over the past two years, DeepSeek has maintained a rapid pace of model releases: V3 launched in late 2024, R1 followed in early 2025, and a series of upgrades—including V3.1 and V3.2—came afterward. In February 2026, some users got an early look through an app update at a gradual rollout that expanded the context window from 128K to 1M, with the knowledge base updated through May 2025. On April 24, the official changelog confirmed the formal launch of DeepSeek-V4.
The official changelog shows that the API now supports two new versions, V4-Pro and V4-Flash, and is compatible with both the OpenAI ChatCompletions API and the Anthropic API. The endpoint remains unchanged; developers only need to set the model parameter to deepseek-v4-pro or deepseek-v4-flash. The old model names deepseek-chat and deepseek-reasoner will be discontinued on July 24, 2026. During the transition period, they will respectively point to V4-Flash’s non-thinking mode and thinking mode. A June report from a third-party cloud platform also said that the production version of V4 introduced a 1M-token context window and a new sparse attention architecture.
A closer look at this update suggests a clear emphasis on efficiency rather than parameter-count showmanship. A previous Nomura Securities report said that two design features introduced in V4—mHC manifold-constrained hyper-connections and the Engram conditional memory module—address two major pain points: instability in deep-network training and high memory usage. The former makes information flow between layers more stable, while the latter places static knowledge in inexpensive memory, freeing up costly HBM. The firm believes V4’s core value lies in using architectural innovations to lower training and inference costs and advance the commercialization of AI applications—not in upending the existing AI value chain. Nor does it expect V4 to recreate the panic over computing demand seen when V3 launched. Notably, the report also mentioned that DeepSeek models accounted for more than half of calls to open-source models on OpenRouter in late 2024, but their share had fallen significantly by the second half of 2025. V4 faces a much more complex competitive environment than it did a year ago.
The impact on crypto markets can only be assessed through logical inference for now; the available evidence contains no data on how related sectors actually responded after V4’s release. There are three possible transmission channels. First, if inference costs fall substantially, the demand narrative for decentralized computing projects could be reassessed, echoing the institutional view that the release will not trigger panic over computing demand. Second, low-cost long-context capabilities could benefit AI agent development: multi-step tasks consume more tokens, potentially providing a catalyst for the agent narrative. Third, as open-source models continue to approach the performance of closed-source models, expectations for local deployment and decentralized inference may strengthen. These are all projections, not established facts.
Editor's note: The clearest signal in this update is the deliberate effort to keep migration costs low. Compatibility with both APIs, combined with a three-month transition period for the old model names, shows that the official priority is the developer ecosystem. For the crypto industry, the ultimate beneficiaries of improved model efficiency are more likely to be application-layer projects that generate real usage and on-chain settlement demand, rather than assets linked to AI only by association. Two sets of data are worth tracking: migration activity after the old model names are discontinued on July 24, and actual changes in application-level usage.
#DeepSeek发布更新版AI模型 #BTC #ETH #BNB