DeepSeek V4 arrives open-weight with 1.6T params and a 7x price gap on Claude
DeepSeek dropped V4 Preview on Friday. Two models, both open-weight, MIT license, weights on Hugging Face the same day. This is the update that was supposed to land in January and got delayed three times. The wait was worth it.
The two models: V4-Pro has 1.6 trillion total parameters with 49 billion active per token. V4-Flash sits at 284 billion total, 13 billion active. Both use Mixture-of-Experts, both have a 1 million token context window with 384K max output. Both text-only. No audio, no video, no images.
That 1.6 trillion figure makes V4-Pro the largest open-weight model available right now, passing Moonshot AI’s Kimi K 2.6 (1.1T) and more than doubling DeepSeek’s own V3.2 (671B).
The numbers that matter. V4-Pro scores 80.6% on SWE-bench Verified. Claude Opus 4.6 sits at 80.8%. That is a 0.2 point gap. The price difference: $3.48 per million output tokens for V4-Pro versus $25 for Claude Opus. Seven times cheaper at near-identical coding benchmark performance. The smaller V4-Flash comes in at $0.14 per million input tokens and $0.28 per million output tokens, undercutting GPT-5.4 Nano, Gemini 3.1 Flash, and Claude Haiku 4.5.
DeepSeek claims V4-Pro-Max outperforms GPT-5.2 and Gemini 3.0 Pro on reasoning benchmarks. Both V4 models score “comparable to GPT-5.4” on coding competition benchmarks. The honest part: DeepSeek admits V4 trails frontier models like GPT-5.4 and Gemini 3.1 Pro on knowledge tests, putting the gap at roughly 3 to 6 months.
How they made 1M context work. This is the technical story that matters. DeepSeek didn’t just slap a bigger context window on and call it done. Three specific architectural changes make million-token inference economically viable instead of a marketing gimmick.
First, a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). The result: at 1M tokens, V4-Pro needs only 27% of the per-token inference FLOPs and 10% of the KV cache compared to V3.2. Ten percent KV cache means you can fit a context that would otherwise require 10x the GPU memory.
Second, Manifold-Constrained Hyper-Connections (mHC). Standard hyper-connections caused training divergence with signal amplification exceeding 3,000x in DeepSeek’s own 27B experiments. The mHC framework constrains residual mixing matrices using the Sinkhorn-Knopp algorithm. Signal amplification drops to 1.6x. The paper was published December 31, 2025, co-authored by CEO Liang Wenfeng. That date tells you this was the core training innovation for V4.
Third, the Muon optimizer replaces AdamW for pre-training. Combined with mHC’s stability guarantees, DeepSeek pushed training to 33 trillion tokens without gradient collapse at 1.6 trillion parameter scale.
The distillation elephant in the room. One day before the V4 launch, the White House Office of Science and Technology Policy issued a memorandum warning about “industrial-scale campaigns” by foreign entities, principally China, to distill US frontier AI models using “tens of thousands of proxy accounts” and jailbreaking techniques. The timing was not subtle.
DeepSeek used what it calls On-Policy Distillation (OPD) for V4, drawing outputs from 10 separate teacher models. The process lets V4 generate its own responses first, then consult multiple teachers to refine them. The White House memo called these “distillation attacks” that “do not replicate the full performance of the original” but allow foreign actors to “release products that appear to perform comparably on select benchmarks at a fraction of the cost.”
The US response includes intelligence sharing with AI companies, cross-industry coordination, developing detection best practices, and exploring accountability measures for foreign actors.
Huawei angle. V4 is specifically adapted for Huawei Ascend chips. This is not an accident. US export controls block DeepSeek from NVIDIA’s latest hardware at scale. The entire architecture, from hybrid attention to mHC, reflects a design philosophy optimized for efficiency over raw compute. The efficiency gains are architectural, not brute-force. Whether that philosophical approach can close the remaining 3-6 month gap without access to frontier training hardware is the open question.
Bottom line for developers. V4-Pro gives you frontier-adjacent coding performance at one-seventh the cost of Claude Opus. V4-Flash gives you a dirt-cheap inference option that still handles complex reasoning. Both are MIT-licensed. Both run at 1M context. Neither handles multimodal. The weights are live on Hugging Face right now.