DeepSeek Rolls Out V4 Preview in Shift to Huawei Chips
DeepSeek launched preview versions of its open-source V4 AI models adapted for Huawei chip technology, marking a shift from Nvidia chips. The release includes V4-Pro, a 1.6 trillion-parameter model with 49 billion active parameters, and V4-Flash, a 284 billion-parameter model with 13 billion active parameters; both are available through the company’s API and support 1 million-token context windows.
DeepSeek said the new architecture uses token-wise compression and DeepSeek Sparse Attention to reduce compute and memory costs, making 1 million-token context the default across its services. The company also said V4-Pro capacity is currently constrained by limited high-end compute and that pricing should fall significantly once Huawei Ascend 950 supernodes are launched at scale in the second half. Artificial Analysis said V4 Pro ranked first among open-weights models on its GDPval-AA benchmark, scoring 1554, ahead of GLM-5.1, MiniMax-M2.7 and Kimi K2.6.
From the sources (25 posts)
@reutersChina's AI darling DeepSeek previews new model adapted for Huawei chip technology
@kimmonismusDeepseek v4 pro Evals. Roughly on par with GPT-5.4 xhigh and opus 4.6 max
@fabknowledgeRT @scaling01: DeepSeek-V4 Benchmarks HuggingFace:
@theahmadosmanDeepSeek V4 IS HERE
@scaling01"DeepSeek-V4-Pro-Max outperforms Opus-4.6-Max on diverse Chinese white-collar tasks, achieving an impressive non-loss rate of 63%"
@zerohedge*DEEPSEEK UNVEILS PREVIEW VERSIONS OF LATEST V4 AI MODEL here we go again
@scaling01DeepSeek-V4 seems to be on a level with Opus 4.5 on real world agentic coding tasks
@mweinbachYea deepseek v4 flash/pro don't really perform that well compared to any of the major US models, even 1-2 revisions old. Looks like it's slightly behind Opus 4.5 in practice, and on par or slightly behind Kimi K2.6 Some good optimization t
@theahmadosmanFINALLY DeepSeek V4 IS HERE - 2 Versions, Flash and Pro - Each comes in Instruct and Base - Flash is 284B MoE with 13B Activated Parameter per Token - Pro is 1.6T MoE w/ 49B Activated Parameter per Token - 1M Context Length The newest
@valsaiDeepSeek v4 is now the #1 open-weight model on our Vibe Code Benchmark, and it’s not close. It leaves the #2 (Kimi K2.6) in the dust, and even beats out frontier closed source models like Gemini 3.1 Pro.
@poezhao0605DeepSeek launched V4 today. Two models, both open-source, both with 1M token context windows. V4-Pro: 1.6T total parameters, 49B active. Benchmarks put it alongside Claude Opus 4.6 and GPT-5.4. V4-Flash: 284B parameters, 13B active. Desig
@yuchenj_uwFinally, DeepSeek V4 is here! - MIT license - DeepSeek-V4-Pro: 1.6T params (49B active) - DeepSeek-V4-Pro Max ≈ Opus-4.6 Max / GPT-5.4 xHigh across benchmarks!
@fabknowledgeRT @scaling01: DEEPSEEK-V4 FLASH AND PRO ITS HAPPENING
@scaling01DeepSeek-V4 has very impressive long context performance and incredibly low cost
@valsaiThe 🐳 has surfaced and it’s a powerhouse on the Vals leaderboards, dominating on coding. DeepSeek V4 just landed #2 on the Vals Index, nearly tying Kimi K2.6 (only 0.07% behind).
@valsaiDeepSeek V4 is text-only with a 1M context window. We evaluated it at temp=1, top_p=0.95, and a max output tokens of either 128k or 256k, depending on the benchmark. We used max reasoning effort for all benchmarks besides terminal bench.
@kimmonismusDeepseek v4 is a huge step upwards compared to DeepSeek 3, outperforms on SWE verified opus 4.6 and GPT-5.4 and sets a new record on Codeforces. Needs to be tested against opus 4.7 and GPT-5.5 tho and see if real world usage holds its prom
@firstadopterDeepSeek V4 is out. Y'all need to stop releasing models in the same week. Seriously. WTH. I'm going to bed.
@emollickAnd now a new DeepSeek model, and appears to be fully open weights. Good benchmarks, but with open models, that isn't always as meaningful. Should be live soon to actually try.
@jukan05Very interesting. DeepSeek added the following comment with V4: “Due to constraints in high-end compute capacity, the current service capacity for Pro is very limited. After the 950 supernodes are launched at scale in the second half of t
@deepseek_ai🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length. 🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models. 🔹 DeepSeek-V4-Fl
@deepseek_aiDeepSeek-V4-Pro 🔹 Enhanced Agentic Capabilities: Open-source SOTA in Agentic Coding benchmarks. 🔹 Rich World Knowledge: Leads all current open models, trailing only Gemini-3.1-Pro. 🔹 World-Class Reasoning: Beats all current open models in
@deepseek_aiDeepSeek-V4-Flash 🔹 Reasoning capabilities closely approach V4-Pro. 🔹 Performs on par with V4-Pro on simple Agent tasks. 🔹 Smaller parameter size, faster response times, and highly cost-effective API pricing. 3/n
@deepseek_aiStructural Innovation & Ultra-High Context Efficiency 🔹 Novel Attention: Token-wise compression + DSA (DeepSeek Sparse Attention). 🔹 Peak Efficiency: World-leading long context with drastically reduced compute & memory costs. 🔹 1M Standard
@deepseek_aiDedicated Optimizations for Agent Capabilities 🔹 DeepSeek-V4 is seamlessly integrated with leading AI agents like Claude Code, OpenClaw & OpenCode. 🔹 Already driving our in-house agentic coding at DeepSeek. The figure below showcases a sa