DeepSeek Releases V4 Under MIT License, Putting a 1.6-Trillion-Parameter Open Model Within Three to Six Months of the Frontier
DeepSeek's V4-Pro and V4-Flash arrive with 1M-token context, three reasoning modes, and pricing that undercuts frontier rivals by up to 8x.
Overview
Chinese AI lab DeepSeek released preview versions of two new large language models on April 24, 2026 — DeepSeek-V4-Pro and DeepSeek-V4-Flash — both published under the MIT License. The release marks the lab’s first major model launch since its R1 reasoning model rattled Silicon Valley in early 2025, and it arrives with a familiar combination: competitive benchmark performance, dramatically lower pricing than Western rivals, and fully open weights available on Hugging Face.
According to DeepSeek’s release announcement, V4-Pro claims to rival the world’s top closed-source models in agentic coding and leads all open-source competitors on those tasks. DeepSeek’s technical report is more measured, noting that performance “falls marginally short of GPT-5.4 and Gemini 3.1-Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately three to six months.”
What We Know
Model specifications
As documented on the Hugging Face model card, V4-Pro is a Mixture-of-Experts model with 1.6 trillion total parameters and 49 billion activated per token — the largest set of open weights released to date. V4-Flash is the efficiency-focused sibling at 284 billion total parameters with 13 billion activated. Both models were pre-trained on more than 32 trillion tokens and support a one-million-token context window.
The architecture introduces several technical innovations over the V3 series, including a Hybrid Attention system combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for efficient long-context handling, and Manifold-Constrained Hyper-Connections (mHC) for improved signal propagation during training. The result is dramatic gains in computational efficiency: at one million tokens, V4-Pro uses only 27 percent of the single-token FLOPs and 10 percent of the KV cache size compared to DeepSeek-V3.2, according to analyst Simon Willison’s breakdown.
Benchmark results
DeepSeek’s self-reported figures for V4-Pro-Max, the highest-reasoning variant, include a LiveCodeBench score of 93.5, a Codeforces rating of 3206, and an SWE-bench Verified score of 80.6 percent on autonomous software engineering tasks. On mathematical reasoning, V4-Pro-Max reaches 95.2 on HMMT 2026 February and 89.8 on IMOAnswerBench. Long-context retrieval at one million tokens scores 83.5 on the MRCR benchmark.
As DataCamp notes, MMLU-Pro comes in at 87.5 and GPQA Diamond at 90.1. Independent evaluations have not yet confirmed these figures.
Reasoning modes
Both models support three operational modes: Non-Think for fast responses on routine tasks, Think High for deliberate logical analysis, and Think Max for maximum reasoning effort. The Think Max mode is designed for open-ended exploration and requires a context window of at least 384,000 tokens to operate effectively, according to the model card.
Pricing
The price gap relative to closed-source competitors is substantial. As reported by Fortune, V4-Pro costs $1.74 per million input tokens and $3.48 per million output tokens. V4-Flash is priced at $0.14 input and $0.28 output. By comparison, DataCamp’s pricing table shows GPT-5.5 at $5.00 input and $30.00 output, and Claude Opus 4.7 at $5.00 input and $25.00 output — making V4-Pro output roughly 8x cheaper than GPT-5.5 and 7x cheaper than Opus 4.7 for the same volume.
Hardware and geopolitical context
The release deepens DeepSeek’s integration with Chinese hardware. As Fortune reports, DeepSeek worked closely with Huawei to ensure V4 runs on the Chinese tech giant’s Ascend AI processors, with Huawei announcing full support for the new models. Fortune also notes that V4’s pricing may fall further as Huawei scales production of its Ascend 950 processors.
The launch comes amid sustained restrictions on DeepSeek’s earlier models: Al Jazeera reports that multiple US states, Australia, Taiwan, South Korea, Denmark, and Italy had all introduced bans or other restrictions on DeepSeek-R1 on privacy and national security grounds. Whether V4 faces similar prohibitions remains to be seen. OpenAI and Anthropic have also previously alleged that DeepSeek employs distillation techniques to leverage capabilities developed at US frontier labs — allegations DeepSeek has not publicly addressed.
Open-source release
Both V4-Pro and V4-Flash are published under the MIT License with full weights available on Hugging Face — 865 GB for V4-Pro and 160 GB for V4-Flash, as noted by Simon Willison. The MIT license imposes no restriction on commercial use or redistribution, making it one of the most permissive licenses applied to a model of this scale. The API, available immediately via deepseek-v4-pro and deepseek-v4-flash model identifiers, is compatible with both the OpenAI ChatCompletions format and the Anthropic Messages API.
What We Don’t Know
Independent benchmark verification has not yet been published. DeepSeek’s self-reported figures have historically been accurate but carry the usual caveats of internally run evaluations. The company has not disclosed details about training infrastructure — specifically whether and to what extent Huawei Ascend chips were used during pre-training versus inference. The scope of DeepSeek’s reported discussions with Tencent and Alibaba, including a reported $40 billion valuation and potential 20 percent stake for Tencent cited by Silicon Republic, remains unconfirmed. And the label “preview” on the release leaves open the question of what changes might come before a full production version is available.
Analysis
DeepSeek’s V4 release follows the same strategic template as its predecessors: open weights, MIT licensing, and pricing that forces a downward reassessment of what large-scale inference should cost. The V4-Flash model at $0.28 per million output tokens now undercuts OpenAI’s own GPT-5.4 Nano ($1.25 output), which was itself positioned as the budget tier.
The benchmark gap with frontier models — three to six months behind, by DeepSeek’s own admission — is smaller than it sounds. DeepSeek-R1 was dismissed as a curiosity until developers ran it in production and found it competitive for most practical tasks. V4-Flash’s SWE-bench Verified score of 79.0 percent would have been a frontier result twelve months ago. For organizations that cannot run closed-source models due to cost, regulatory concerns, or the need for on-premise deployment, V4 significantly raises the ceiling of what is achievable with open weights.
The Huawei Ascend compatibility is a structural development beyond the model itself. It validates a compute path for Chinese AI labs that does not depend on NVIDIA hardware, and it establishes Ascend as a credible inference platform at least for the models Huawei and DeepSeek co-develop. Whether that translates into competitive training infrastructure at the frontier scale remains the key open question.