--- title: DeepSeek V4 launches on Huawei chips, raising questions... description: DeepSeek has released preview versions of V4-Pro (1.6T parameter MoE, 49B active) and V4-Flash (284B total, 13B active) under MIT license on Hugging Face. Both... canonical: https://daily.dev/posts/deepseek-v4-launches-on-huawei-chips-raising-questions-about-cuda-s-grip-on-ai-lxuventxc twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: DeepSeek V4 launches on Huawei chips, raising questions about CUDA's grip on AI | daily.dev og:description: DeepSeek has released preview versions of V4-Pro (1.6T parameter MoE, 49B active) and V4-Flash (284B total, 13B active) under MIT license on Hugging Face. Both... og:url: https://daily.dev/posts/deepseek-v4-launches-on-huawei-chips-raising-questions-about-cuda-s-grip-on-ai-lxuventxc og:image: https://api.daily.dev/og/posts/lXUVeNTxC.png og:image:alt: DeepSeek V4 launches on Huawei chips, raising questions about CUDA's grip on AI og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # DeepSeek V4 launches on Huawei chips, raising questions about CUDA's grip on AI **[Collections](https://daily.dev/sources/collections)** · 4 min read · 2 upvotes · 0 comments ## Summary DeepSeek has released preview versions of V4-Pro (1.6T parameter MoE, 49B active) and V4-Flash (284B total, 13B active) under MIT license on Hugging Face. Both support a 1M token context window via a new hybrid attention architecture (CSA+HCA) that cuts inference FLOPs by 73% and KV cache memory by 90% vs V3.2. Pricing is aggressive — V4-Pro at ~$1.74/M input tokens at full price, with a 75% promo discount until May 2026 — though competitors like Qwen3 and Kimi K2 are now in the same range. V4-Flash can run locally on consumer hardware with selective quantization at 30–34 tokens/second on an M3 Ultra. The most strategically significant detail: DeepSeek optimized V4 for Huawei Ascend 950 and Cambricon chips, migrating inference away from CUDA/Nvidia. Jensen Huang called this a 'horrible outcome' for the US, noting CUDA dominance has served as a layer of American control over AI. Export controls intended to slow China's AI progress may instead be accelerating domestic Chinese AI hardware development. ## Content # DeepSeek V4: what's actually new, what it costs, and why the Huawei chip angle matters DeepSeek has released preview versions of two new models — V4-Pro and V4-Flash — on Hugging Face under an MIT license. Both are open-weight and available via API today. ## The two models **V4-Pro** is a 1.6 trillion parameter mixture-of-experts model with 49 billion active parameters. It targets complex coding, multi-step reasoning, and agentic tasks, and claims the top spot among open-source models on coding and math benchmarks. DeepSeek acknowledges it trails closed-source frontier models like GPT-5.4 and Gemini 3.1 Pro by roughly 3–6 months. **V4-Flash** is the smaller sibling: 284 billion total parameters, 13 billion active. It's faster and cheaper, aimed at tasks where you don't need the full Pro capability. Both models default to a 1 million token context window, enabled by a new hybrid attention architecture that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). DeepSeek claims this cuts per-token inference FLOPs by 73% and KV cache memory by 90% compared to V3.2 — which is what makes the 1M context window practical rather than theoretical. The existing `deepseek-chat` and `deepseek-reasoner` API endpoints will be retired on July 24, 2026. ## Pricing This is where things get genuinely strange. DeepSeek is running a 75% promotional discount on V4-Pro until May 5, 2026, dropping input prices to around $0.036 per million tokens. Even at full price — roughly $1.74 per million input tokens and $0.87 per million output tokens — it undercuts OpenAI GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.7 significantly. Opus 4.7 runs $25 per million output tokens. That's about a 28x gap. DeepSeek also permanently cut cache-hit prices to one-tenth of previous levels across its entire API suite. To put the output pricing in concrete terms: $5 gets you roughly 5.75 million output tokens, or about 14,000 pages of text. That said, V4-Pro's pricing is noticeably higher than V3's debut pricing, and the competitive landscape has shifted. Qwen3, Kimi K2, and Xiaomi's MiMo-V2 Pro are all in the same cost range now. "Cheap and good" is no longer a DeepSeek exclusive. ## Running it locally V4-Flash is small enough to run on consumer hardware with aggressive quantization. Developer antirez has been running it locally using a selective 2-bit quantization scheme: routed experts get quantized hard (IQ2_XXS for input/gate, Q2_K for output), while shared experts, projection layers, and routing weights stay at Q8 to preserve quality. On an M3 Ultra, this runs at around 30–34 tokens per second and handles tool calling well enough to power parallel agent workflows. His take after 24 hours: "it is the FIRST time I feel I have a frontier model running on my computer." NVIDIA is also offering V4-Pro and V4-Flash through its NIM APIs with free developer access for prototyping. The endpoints are OpenAI-compatible and support a `reasoning_effort` parameter (none/high/max). On GB200 NVL72 hardware, V4-Pro benchmarks at over 150 tokens per second per user. ## The Huawei chip angle This is probably the most significant part of the release that isn't getting enough attention. DeepSeek optimized V4 for Huawei's Ascend 950 chips and Cambricon hardware, not Nvidia or AMD. That means migrating from CUDA to Huawei's CANN framework. The inference workload is being shifted to Chinese domestic chips, with prices expected to drop further once Huawei's Ascend 950 supernodes ship at scale. DeepSeek still appears to rely on Nvidia for training, but the inference side is moving. Nvidia CEO Jensen Huang called this a "horrible outcome" for the US in a recent podcast appearance. His concern isn't just market share — it's that CUDA's dominance has functioned as a second layer of American control over AI development. If DeepSeek can run inference efficiently on Huawei chips, that dependency weakens. Huawei's Ascend currently delivers roughly 60% of H100 inference performance, and it reportedly couldn't handle DeepSeek's R2 training workloads reliably. But Huang's point is that software optimization and researcher focus can close hardware gaps over time. US lawmakers are pushing to add DeepSeek to the export control entity list. There's an uncomfortable irony here: the export controls that were meant to slow China's AI development may be accelerating the build-out of a domestic Chinese AI hardware ecosystem. The V4 release lands in a more competitive market than R1 did a year ago — OpenAI shipped GPT-5.5 the same day, Anthropic is valued at $1 trillion, and Google released Gemma 4 with multimodal support and local deployment. V4 is a genuinely strong model. It's just no longer arriving in a vacuum. --- Tags: [#llm](https://daily.dev/tags/llm), [#cuda](https://daily.dev/tags/cuda), [#deepseek](https://daily.dev/tags/deepseek) [View this post on daily.dev](https://daily.dev/posts/deepseek-v4-launches-on-huawei-chips-raising-questions-about-cuda-s-grip-on-ai-lxuventxc) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"DeepSeek V4 launches on Huawei chips, raising questions about CUDA's grip on AI","url":"https://daily.dev/posts/deepseek-v4-launches-on-huawei-chips-raising-questions-about-cuda-s-grip-on-ai-lxuventxc","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/deepseek-v4-launches-on-huawei-chips-raising-questions-about-cuda-s-grip-on-ai-lxuventxc"},"datePublished":"2026-04-24T13:30:05.079Z","dateModified":"2026-04-28T14:27:22.146Z","description":"DeepSeek has released preview versions of V4-Pro (1.6T parameter MoE, 49B active) and V4-Flash (284B total, 13B active) under MIT license on Hugging Face. Both...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/deepseek-v4-launches-on-huawei-chips-raising-questions-about-cuda-s-grip-on-ai-lxuventxc","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,cuda,deepseek","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"DeepSeek V4 launches on Huawei chips, raising questions about CUDA's grip on AI"}]} ```