On April 24, 2026, DeepSeek released preview versions of its V4 family, V4-Pro and V4-Flash, under an MIT open-weights license hosted on Hugging Face.1 TechCrunch 2026-04-24 V4-Pro and V4-Flash launched April 24, 2026 with 1M token context; reasoning trails GPT-5.4 and Gemini 3.1 Pro by an estimated three to six months, coding comparable to GPT-5.4. Open source 2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source V4-Pro uses a Mixture-of-Experts design with 1.6 trillion total parameters and 49 billion active per token, while V4-Flash runs 284 billion total and 13 billion active, and both carry a 1 million token context window, up from the 128,000 tokens of V3.1 TechCrunch 2026-04-24 V4-Pro and V4-Flash launched April 24, 2026 with 1M token context; reasoning trails GPT-5.4 and Gemini 3.1 Pro by an estimated three to six months, coding comparable to GPT-5.4. Open source 3 TechXplore 2026-04-24 V4 upgrades context to 1M tokens from V3's 128k, claims improved reasoning and agentic ability, and Huawei Ascend chips are now compatible with the models. Open source The stake is not a new capability ceiling. It is price and access: DeepSeek prices V4-Pro output at $3.48 per million tokens against roughly $25 to $30 for the comparable tier from OpenAI and Anthropic.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source We assess with moderate confidence that V4 does not lead the frontier but does compress its economics, and that the durable effect is on the cost curve for capable open models rather than on the capability record.

What the release actually shows

The measured gap is now stated in months. On reasoning benchmarks DeepSeek claims V4 has nearly closed the distance to leading models but falls slightly behind on knowledge tests, a lag that reporting placed at roughly three to six months behind GPT-5.4 and Gemini 3.1 Pro.1 TechCrunch 2026-04-24 V4-Pro and V4-Flash launched April 24, 2026 with 1M token context; reasoning trails GPT-5.4 and Gemini 3.1 Pro by an estimated three to six months, coding comparable to GPT-5.4. Open source 2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source On coding the picture is stronger, with performance described as comparable to GPT-5.4.1 TechCrunch 2026-04-24 V4-Pro and V4-Flash launched April 24, 2026 with 1M token context; reasoning trails GPT-5.4 and Gemini 3.1 Pro by an estimated three to six months, coding comparable to GPT-5.4. Open source On agentic tasks DeepSeek's internal evaluations put V4-Pro near Claude Opus 4.5, though that placement rests on the company's own numbers and is not independently verified.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source 3 TechXplore 2026-04-24 V4 upgrades context to 1M tokens from V3's 128k, claims improved reasoning and agentic ability, and Huawei Ascend chips are now compatible with the models. Open source That caveat matters: benchmark self-reports from labs shipping their own weights should be read as marketing until third parties reproduce them.

The release also ends a long public silence. V4 is DeepSeek's first major model in 484 days, since V3.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source The company paired it with efficiency claims, a 73% reduction in per-token inference FLOPs and a 90% cut in KV cache memory versus its V3.2 line, again unverified by outside researchers.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source Even discounted, the direction is clear. The pitch is not that V4 is smarter than the frontier. It is that near-frontier reasoning and frontier-class coding are now available as open weights at a fraction of the token cost.

The compute story underneath

The most consequential detail sits in the hardware. V4 was trained on Nvidia silicon but runs inference on both Nvidia Blackwell endpoints and Huawei Ascend processors, and DeepSeek says Ascend chips are now compatible with the models.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source 3 TechXplore 2026-04-24 V4 upgrades context to 1M tokens from V3's 128k, claims improved reasoning and agentic ability, and Huawei Ascend chips are now compatible with the models. Open source This is a split-path architecture: US chips still govern the training frontier, but the serving layer, where the bulk of lifetime compute demand lives, can route around export controls onto domestic Chinese hardware.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source We assess with moderate confidence that this is the more strategically important fact than any benchmark line, because it converts export-control pressure from a hard ceiling into a manageable cost on one half of the pipeline.

Who gains and who loses follows from that split. The clearest winner is Huawei, whose Ascend line gains a flagship reference workload that demonstrates practical inference at scale outside Nvidia's ecosystem.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source 3 TechXplore 2026-04-24 V4 upgrades context to 1M tokens from V3's 128k, claims improved reasoning and agentic ability, and Huawei Ascend chips are now compatible with the models. Open source Cost-sensitive enterprises and developers gain an open model priced roughly a tenth of the closed Western tier on output tokens, which they can host themselves.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source The clearest losers are the closed-model providers whose pricing power at the mid-capability tier erodes every time an open model reaches good-enough quality, and Nvidia at the margin, which keeps the training business but watches inference demand partially migrate to Ascend inside China.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source DeepSeek itself is a more ambiguous case. It no longer stands alone as the Chinese open flagship: Alibaba's Qwen3, Moonshot's Kimi, Zhipu's GLM, ByteDance's Doubao, and MiniMax are shipping on similar tiers, which fragments rather than consolidates the field it once defined.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source

The counter-case

The thesis that V4 meaningfully shifts the market could be wrong in two ways. First, the efficiency and agentic claims are DeepSeek's own, and if independent testing shows the reasoning gap is closer to a full generation than to three months, the price advantage matters less because buyers of frontier work will not trade capability for cost.1 TechCrunch 2026-04-24 V4-Pro and V4-Flash launched April 24, 2026 with 1M token context; reasoning trails GPT-5.4 and Gemini 3.1 Pro by an estimated three to six months, coding comparable to GPT-5.4. Open source 2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source Second, open weights at low prices do not automatically win enterprise deployment, where data governance, support, liability, and the reluctance of Western buyers to run a Chinese model on sensitive workloads all cut against adoption regardless of the benchmark. V4 is also text-only, with no native audio, video, or image handling, which limits it against multimodal rivals for a growing share of use cases.1 TechCrunch 2026-04-24 V4-Pro and V4-Flash launched April 24, 2026 with 1M token context; reasoning trails GPT-5.4 and Gemini 3.1 Pro by an estimated three to six months, coding comparable to GPT-5.4. Open source For the thesis to fail, the capability gap would have to be wider than claimed, or the non-price frictions would have to keep V4 confined to hobbyist and domestic-Chinese use.

What to watch

  • Independent benchmarks land. If third-party evaluations over the next one to two quarters confirm V4-Pro within three to six months of GPT-5.4 on reasoning and at parity on coding, the price story holds; a wider verified gap would reprice it as a value model, not a frontier-adjacent one.1 TechCrunch 2026-04-24 V4-Pro and V4-Flash launched April 24, 2026 with 1M token context; reasoning trails GPT-5.4 and Gemini 3.1 Pro by an estimated three to six months, coding comparable to GPT-5.4. Open source
  • Ascend inference share. Watch for evidence that V4 inference volume on Huawei Ascend grows through 2026. Visible enterprise deployments serving V4 on Ascend would confirm the export-control workaround is real capacity, not a demo.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source 3 TechXplore 2026-04-24 V4 upgrades context to 1M tokens from V3's 128k, claims improved reasoning and agentic ability, and Huawei Ascend chips are now compatible with the models. Open source
  • Western closed pricing responds. If OpenAI or Anthropic cut mid-tier token prices within two quarters, treat it as direct evidence that open models at V4's price are exerting real competitive pressure on the closed frontier.2 WinBuzzer 2026-04-27 MIT open weights, 1.6T total / 49B active MoE, first major release in 484 days, priced far below GPT-5.5 and Opus 4.7, inference on Nvidia Blackwell and Huawei Ascend, with unverified efficiency claims. Open source
  • Enterprise trust barrier. Track whether any large non-Chinese enterprise publicly standardizes on V4 for production workloads by year end; absence would confirm that governance and provenance, not capability, are the binding constraint on adoption.

The signal to carry forward is that the frontier race and the deployment race are separating. V4 is unlikely to top a leaderboard, but by pushing capable open weights onto domestic silicon at a tenth of the price, it makes the question for most buyers less about who leads and more about what near-frontier work now costs.