DeepSeek says V4 matches GPT-5.4 on standard benchmarks
- DeepSeek’s April 24 V4 release is real, but the official materials do not say it “matches GPT-5.4” across standard benchmarks in general. - What DeepSeek actually published: V4-Pro at 1.6T parameters with 49B active, V4-Flash at 285B with 13B active, both at 1M context. - That matters because the real story is efficiency and open weights — not a clean frontier-model tie claim DeepSeek itself clearly documents.
Large language models are the domain here, but the actual story is narrower than the headline makes it sound. DeepSeek did release V4 on April 24, 2026, with open weights, API access, and a technical report. But the clean claim that “V4 matches GPT-5.4 on standard benchmarks” is not something the official release plainly spells out. The gap is between what people are repeating on social media and what DeepSeek itself actually published. ### What did DeepSeek actually launch? DeepSeek launched a family, not one single model. The official docs list DeepSeek-V4-Pro and DeepSeek-V4-Flash, both released on April 24, 2026, and both available through open repositories and the API. The API changelog shows the same date and says the older `deepseek-chat` and `deepseek-reasoner` names are being phased out in favor of V4 model names. ### What are the concrete model sizes? This is one place where the official numbers are very clear. DeepSeek says V4-Pro has 1.6T total parameters with 49B activated per token, while V4-Flash has 285B total parameters with 13B activated per token. Both are Mixture-of-Experts models with a 1M-token context window. So if you saw “284B with 13B active,” that is basically the Flash model, not the whole V4 line. ### Is V4 multimodal? Not from the official model card we can verify. The published technical documentation describes V4 as supporting text, with three reasoning modes — Non-think, Think High, and Think Max. The model card does not present V4 as a general multimodal release in the way the social summary suggests. So what's the real technical angle? The technical angle is long-context efficiency. DeepSeek says V4 keeps the DeepSeekMoE setup and Multi-Token Prediction from V3, then adds a hybrid attention design combining CSA and HCA, plus manifold-constrained hyper-connections and the Muon optimizer. In plain English, this is about making bigger. ### So where does the GPT-5.4 comparison come from? Mostly from secondary writeups and comparison blogs, not from the clearest official phrasing. DeepSeek’s own preview page says V4-Pro has “world-class reasoning,” beats current open models in math, STEM, and coding, and rivals top closed-source models. That is a strong claim, but it is softer than “matches GPT-5.4 on standard benchmarks” across the board. ### What about post-training and distillation? That part does appear in the Hugging Face model page. DeepSeek describes a two-stage post-training setup: first, domain-specific experts are cultivated with supervised fine-tuning and RL using GRPO; then those capabilities are merged through on-policy distillation into one unified model. That matters because it stems from aggressive post-training. ### Why are people focusing on cost? Because DeepSeek is pushing V4 as a cheaper frontier-style option. The pricing page shows a temporary 75% discount for `deepseek-v4-pro` through May 31, 2026, and reduced cache-hit pricing from April 26. Open weights plus lower pricing is the competitive wedge here. The point is not merely “we tied OpenAI.” The point is “we got close enough, and we can distribute it differently.” ### Bottom line? DeepSeek did publish a serious V4 release. The verified story is a 1M-context MoE family with open weights, heavy efficiency work, and strong claims against top closed models. But the exact “matches GPT-5.4 on standard benchmarks” line looks overstated unless you rely on secondary summaries rather than DeepSeek’s own published materials.
