On April 24, 2026, DeepSeek dropped preview versions of its V4 series — and it's the biggest open-source model release so far this year. Two models, both Mixture-of-Experts, both supporting a native one-million-token context window, both open weights under MIT license. The headline: DeepSeek V4 delivers frontier-level performance at prices that undercut every major competitor.
Two Models, Two Targets
| Model | Total Params | Active Params | Hugging Face Size | Pricing (In/Out per M tokens) |
|---|---|---|---|---|
| DeepSeek-V4-Flash | 284B | 13B | 160GB | $0.14 / $0.28 |
| DeepSeek-V4-Pro | 1.6T | 49B | 865GB | $1.74 / $3.48 |
V4-Flash is the cost-effective default for everyday tasks. V4-Pro is the heavy hitter, aimed at reasoning, coding, and complex agentic workflows where maximum intelligence matters. Both support Thinking and Non-Thinking modes — letting you trade speed for accuracy depending on the task.
How Cheap Is It, Really?
DeepSeek V4 Flash at $0.14/$0.28 per million tokens is cheaper than GPT-5.4 Nano ($0.20/$1.25). DeepSeek V4 Pro at $1.74/$3.48 undercuts GPT-5.4 ($2.50/$15.00), Claude Sonnet 4.6 ($3.00/$15.00), and Gemini 3.1 Pro ($2.00/$12.00) by significant margins. For teams running high-volume inference, the savings are dramatic.
For context, here's how DeepSeek V4 Pro stacks up against current frontier models on pricing:
- GPT-5.5: $5.00 / $30.00 per million tokens
- Claude Opus 4.7: $5.00 / $25.00
- GPT-5.4: $2.50 / $15.00
- DeepSeek V4 Pro: $1.74 / $3.48
- DeepSeek V4 Flash: $0.14 / $0.28
The Pro model delivers roughly GPT-5.2 to GPT-5.4 level performance at less than a quarter of the cost.
The Architecture Behind the Numbers
Three key innovations power the V4 release:
1. Hybrid CSA and HCA Attention. Standard transformer attention gets prohibitively expensive at one million tokens. V4 introduces Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to split the problem. CSA compresses KV caches along the sequence dimension, then applies sparse attention with a learned indexer. The result: at 1M tokens, V4-Pro uses only 27% of the FLOPs and 10% of the KV cache of V3.2. V4-Flash drops to 10% FLOPs and 7% cache.
2. Manifold-Constrained Hyper-Connections (mHC). An upgrade to residual connections designed for numerical stability in very deep stacks — the kind of thing that matters when you're pushing past 60 layers at trillion-parameter scale.
3. Muon Optimizer. V4 switches from AdamW to Muon for most parameters during training, reporting faster convergence and more stable optimization at scale.
Both models also use FP4 storage for routed expert weights (halving memory vs FP8), Multi-Token Prediction (inherited from V3), and a 128K vocabulary tokenizer.
Benchmark Performance
DeepSeek reports V4-Pro competitive with current frontier models, with a notable caveat: it sits between GPT-5.2 and GPT-5.4 on most reasoning benchmarks, trailing the state of the art by approximately 3–6 months. On Codeforces, V4-Pro achieves a 2,206 rating, ranking 23rd among human competitors.
The honest framing from DeepSeek's own paper: the model doesn't beat GPT-5.4 or Gemini 3.1 Pro outright. What it does is get within striking distance at a fraction of the price — and do it with fully open weights.
Running Locally
V4-Flash at 160GB means a lightly quantized version could plausibly run on a 128GB M5 MacBook Pro. V4-Pro at 865GB is more challenging but may work with expert streaming from disk. The Unsloth team is expected to release quantized versions soon.
Both models are available now on Hugging Face, through the DeepSeek API (supporting both OpenAI and Anthropic API formats), and at chat.deepseek.com.
What This Means
DeepSeek V4 isn't the smartest model in the room — GPT-5.5, released one day earlier on April 23, holds that crown. But V4 changes the economics of frontier-level AI. When you can get 90%+ of the performance at 20% of the cost, with open weights you can self-host, the "just use the best model" calculus shifts significantly.
For developers and teams building AI products, V4-Flash is the new go-to for high-volume tasks. V4-Pro is the budget frontier model. And both push the industry toward a future where frontier-level intelligence isn't locked behind premium pricing.
Key Specs Summary
- License: MIT (open weights)
- Context window: 1 million tokens (both models)
- Training data: 32–33 trillion tokens
- MoE routing: Hash routing (first 3 layers) + learned routing (remaining layers)
- Precision: FP8 weights, FP4 expert weights
- Availability: Hugging Face, DeepSeek API, chat.deepseek.com
- API compatibility: OpenAI ChatCompletions + Anthropic formats
- V3/V3.2 retirement: July 24, 2026