One year ago, DeepSeek dropped R1. Markets reeled. Nvidia shed nearly $600 billion in market cap as investors questioned Big Tech’s AI spending frenzy. The Hangzhou-based startup claimed it built a reasoning powerhouse for under $6 million using lower-end Nvidia H800 chips—90-95% cheaper than rivals. CNBC chronicled the fallout. Now, on April 26, 2026, DeepSeek strikes again with V4 preview. Open weights. 1 million token context as standard. Costs slashed further.
R1 wasn’t hype. A 671 billion parameter Mixture-of-Experts model with 37 billion active parameters and 128K context. Trained on DeepSeek-V3-Base via reinforcement learning—no supervised fine-tuning for the Zero variant. Pure RL unlocked self-verification, reflection, long chain-of-thought. But repetition plagued it. R1 fixed that with cold-start data and dual RL-SFT pipeline.
Benchmarks stunned. MATH-500: 97.3% pass@1, topping OpenAI o1-1217’s 96.4. AIME 2024: 79.8 versus 79.2. Codeforces rating: 2029 against o1’s 2061. GPQA-Diamond: 71.5 to o1’s 75.7. MMLU-Pro: 84.0, best listed. Hugging Face model card. Distilled versions—Qwen-32B, Llama-70B—beat o1-mini on AIME (72.6%), MATH-500 (94.5%). MIT licensed. Six small models free for commercial use. API at $0.14/million input (cache hit), $2.19 output. DeepSeek API Docs.
App stores bowed. By January 27, 2025, DeepSeek-R1 topped U.S. iOS free apps, eclipsing ChatGPT. Wikipedia. Bans followed in U.S. states, Australia, Taiwan—privacy fears. Yet adoption surged. Available on Amazon Bedrock, SageMaker JumpStart, GitHub Models, Snowflake Cortex. AWS blog.
R1’s Lasting Ripples in Reasoning and Deployment
And then silence. DeepSeek iterated quietly—R1-0528 in May 2025 cut false outputs, boosted complex tasks toward o3, Gemini 2.5 Pro. Reuters. V3.1-Think sped reasoning. But V4? A leap. Pro: 1.6 trillion total parameters, 49 billion active. Flash: 284 billion total, 13 billion active. Both 1M context native.
Token-wise compression. DeepSeek Sparse Attention. Compute down 73%, KV cache memory to 10% of V3.2 for 1M tokens. Pretrained on 32 trillion tokens. Two-stage post-training: domain experts, then distillation. mHC connections, Muon optimizer for trillion-scale stability. Agent-optimized—pairs with Claude Code, OpenClaw. DeepSeek announcement.
V4-Pro leads open models in agentic coding—SOTA benchmarks. World knowledge trails only Gemini-3.1-Pro. Tops opens in math/STEM/coding, rivals closed leaders. Flash nears Pro on simple agents, faster, cheaper. API live: deepseek-v4-pro, -flash. Thinking/Non-Thinking modes. Old chat, reasoner retire July 24. Weights on Hugging Face. Hugging Face collection.
Neil Shah of Counterpoint called V4 a “serious flex” for agent tasks at low cost. CNBC. TechCrunch notes V4-Pro-Max outstrips GPT-5.2, Gemini 3.0 on reasoning; coding matches GPT-5.4. Lags knowledge by 3-6 months. TechCrunch. WSJ: Matches late-2025 U.S. tiers, trails Claude Opus 4.6, Gemini 3.1 Pro in spots. Output tokens: $3.50/million versus Anthropic’s $25. Wall Street Journal.
V4 Ushers Cost-Effective Frontiers—and Market Tremors
So what now? R1 proved reasoning via RL scales cheap. V4 scales context cheap. Open labs narrow U.S. lead to months, per Stanford AI Index. Huawei Ascend compatibility eyes China’s chip curbs. Developers grab 1M prompts for repo agents, long planners. Small teams self-host frontier inference.
Markets watch. R1 erased $1 trillion in tech/energy stocks. V4 reignites debate: Does efficiency kill capex arms race? DeepSeek bets yes. Compute migrates to gradients, not tokens. Org charts lag. Budgets shift.
Frontier closed models—GPT-5.5, Claude Mythos—push agents, science. But DeepSeek open-weights match enough, cost less. R1 app topped charts. V4 agents deploy now. The gap? Closing fast.
WebProNews is an iEntry Publication