What DeepSeek V4 Brings to the Table
DeepSeek V4 just dropped under the MIT license with a 1-million-token context window and pricing that undercuts every major provider by an order of magnitude. The benchmark numbers look impressive, with the Pro model competing against GPT-5.4 and Claude Opus 4.8 on several standard tests. After putting both models through front-end builds, agentic workflows, and coding challenges, though, the gap between leaderboard scores and real practical output is wider than you'd expect.
The release includes two distinct models built on a new architecture, DeepSeek's first major structural change since V3. Both support hybrid thinking and non-thinking modes, and both ship with open weights on HuggingFace and Ollama Cloud. Here's what you need to know before integrating either one into your stack.
DeepSeek V4 Specs: Pro vs. Flash
DeepSeek's two-tier approach gives developers a clear choice between raw power and cost efficiency. Both models share the same 1M-token context window and 384K max output, but they differ significantly in size, speed, and price.
DeepSeek V4 Pro (deepseek-v4-pro)
The flagship runs 1.6 trillion total parameters with 49 billion active at any given time, thanks to its Mixture-of-Experts design. Pricing sits at $1.74 per million input tokens on a cache miss, dropping to $0.145 on a cache hit. Output costs $3.48 per million tokens. For comparison, Claude Opus 4.8 charges roughly $15/$75 for input/output per million tokens. Pro's pricing makes it roughly 20x cheaper on output alone.
DeepSeek V4 Flash (deepseek-v4-flash)
Flash is the lightweight counterpart at 284 billion total parameters and 13 billion active. At $0.14 per million input tokens (cache miss) and $0.28 per million output tokens, it costs roughly 10x less than Pro across the board. That pricing puts it in a category of its own for high-volume, cost-sensitive applications. If you're building something where request volume matters more than peak reasoning quality, Flash is the clear starting point.
Migration heads-up: The legacy deepseek-chat and deepseek-reasoner model IDs will be deprecated on July 24, 2026, just three days from now. They currently map to V4 Flash's non-thinking and thinking modes. If you're using those endpoints, switch to deepseek-v4-flash or deepseek-v4-pro immediately. Our machine learning course covers model migration patterns if you need a structured approach.
Architecture Changes in DeepSeek V4
V4 introduces several technical advances over DeepSeek-V3.2. These aren't minor tweaks; the structural changes directly affect inference costs and long-context performance.
Hybrid Attention Architecture combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). At the full 1M-token context, V4 Pro requires only 27% of the single-token inference FLOPs and 10% of the KV cache compared to V3.2. For anyone running long-context production workloads, those are meaningful efficiency gains.
Manifold-Constrained Hyper-Connections (mHC) stabilize signal propagation across the model's deep layer stack without limiting expressivity. The Muon Optimizer delivers faster convergence during training, which ran on a dataset exceeding 32 trillion tokens. Mixed Precision Training uses FP4 for MoE expert parameters and FP8 for most other parameters, balancing memory efficiency against output quality.
The result is a model that handles massive context windows at a fraction of the compute cost you'd need for a dense architecture of similar quality.
DeepSeek V4 Benchmarks: The Numbers
On standard academic benchmarks, V4 Pro consistently outperforms V3.2. MMLU rises from 87.8 to 90.1, HumanEval Pass@1 jumps from 62.8 to 76.8, and MMLU-Pro climbs from 65.5 to 73.5. Flash also improves over V3.2 in most areas, though by smaller margins.
In maximum reasoning mode, V4 Pro competes directly with closed-source leaders:
- LiveCodeBench Pass@1: 93.5, ahead of Gemini-3.1-Pro High (91.7) and Opus-4.6 Max (88.8)
- GPQA Diamond Pass@1: 90.1, behind GPT-5.4 xHigh (93.0) and Gemini-3.1-Pro High (94.3)
- SWE Verified Resolved: 80.6, matching Gemini-3.1-Pro High
- Codeforces Rating: 3206 in competitive programming, ahead of Gemini-3.1-Pro High
On the GDPval-AA agentic benchmark, V4 Pro leads open-weights models with a score of 1554, ahead of GLM-5.1 (1535), MiniMax-M2.7 (1514), and Kimi K2.6 (1484). The Artificial Analysis Intelligence Index ranks V4 Pro at #2 among open-weights reasoning models, trailing only Kimi K2.6.
Those numbers are real. But benchmarks and real-world output measure different things.
Real-World Testing: Where DeepSeek V4 Falls Short
Practical testing reveals a pattern: the models feel "benchmark maxed," optimized to ace standardized tests rather than produce polished, usable output. On the Code Arena leaderboard, V4 currently sits at #3, behind GLM 5.1 and Kim K 2.6.
Here's how V4 performed across a gauntlet of front-end, coding, and agentic tasks:
- Mac OS / Slack clone: Lackluster and basic, with no SVG icons. Lost to GLM 5.1.
- SVG butterfly generation: Easily beaten by Qwen 3.6 Plus.
- 3D PS5 controller (Three.js): The output looked like a table. Qwen 3.6 Plus won.
- EV durability test / Instagram clone: Bugged out and failed. Minimax M2.7 handled it.
- SaaS landing page: Output quality felt like GPT-3.5 era generation.
- Minecraft clone: Pro produced a messy sandbox with no lighting. Flash generated infinite terrain and a working inventory system.
- 360° product viewer: Solid 6/10 from Pro.
The Minecraft result deserves attention. The cheaper Flash model outperformed the flagship Pro on a complex task, generating infinite terrain and inventory management while Pro got stuck on basic sandbox functionality. For developers evaluating both models, this matters: the more expensive option isn't always the better one.
Comparing DeepSeek V4 to a top-tier model like Claude Opus 4.8 isn't even fair right now. Opus produces significantly more polished output across the board. Models like Qwen 3.6 Plus, Minimax M2.7, and GLM 5.1 also outperform V4 in extended thinking and real-world scenarios at the moment.
The Hallucination Problem
One specific limitation deserves attention. On the AA-Omniscience benchmark, V4 Pro has a 94% hallucination rate. When the model doesn't know an answer, it fabricates one 94% of the time instead of acknowledging uncertainty.
For applications where confidence calibration matters, this is a serious concern. If you're building anything where incorrect answers carry consequences, medical information, financial guidance, legal research, you'll need guardrails beyond the model's default behavior. Our AI fundamentals course covers practical approaches to hallucination detection and mitigation in production systems.
Configurable Reasoning Depth
One of V4's more useful features is per-request reasoning control. You pick the level of thinking effort depending on the task:
- Non-think: Fast, low-latency responses for routine tasks, simple chat, and low-risk decisions.
- Think High: Logical analysis with moderate latency. Good for complex problem-solving, planning, and coding.
- Think Max: Maximum reasoning depth. Use this for the hardest problems where accuracy matters more than speed.
This gives you fine-grained control over the cost-latency-accuracy tradeoff. Run simple classification in non-think mode at Flash pricing, then switch to Pro in think-max mode for architecture decisions or algorithm design. The ability to adjust per-request is especially useful when your application handles both lightweight and complex queries in the same pipeline.
Getting Started with the DeepSeek V4 API
The API is available at api.deepseek.com with both OpenAI-compatible and Anthropic-compatible endpoints. Supported features include JSON mode, tool calls, fill-in-the-middle (FIM) completion, and chat-prefix completion.
Open weights live on HuggingFace and Ollama Cloud for local deployment. You can also access the models for free through the official DeepSeek chatbot. For developers working in Python, the API integration follows the same patterns as any OpenAI-compatible service, which means existing codebases need minimal changes to switch over.
If you're connecting DeepSeek V4 to your development environment, here's a quick setup walkthrough:
Our Python programming course covers API integration patterns that work with any OpenAI-compatible endpoint, including DeepSeek's, if you want a more thorough grounding.
Pros and Cons at a Glance
What DeepSeek V4 Does Well
- Massive 1M context window with efficient inference
- MIT license means fully open weights, no restrictions
- Extremely low pricing compared to closed-source alternatives
- Flash model punches above its weight class on specific tasks
- Strong on 3D rotation and spatial reasoning benchmarks
- Configurable reasoning depth per request
Where It Needs Work
- Output feels "benchmark maxed" in real-world use
- Sloppy execution on front-end and agentic tasks
- Lags behind GLM 5.1, Qwen 3.6 Plus, and Kimi K2.6 on extended thinking
- 94% hallucination rate on unknown-answer tasks
- Pro model fails complex agentic flows that competitors handle
Recommended Courses
Keep learning with these free courses:
Should You Use DeepSeek V4 Right Now?
The answer depends on what you're building.
Good fit: High-volume applications where cost is the primary constraint. Long-context retrieval tasks where the 1M window provides real value. Prototyping and experimentation where Flash's near-zero pricing makes rapid iteration affordable. Tasks involving 3D rotation or spatial reasoning, where V4 performs above its weight class.
Not ready yet: Front-end code generation where output polish matters. Complex agentic workflows with multiple decision points. Applications requiring calibrated confidence scores or low hallucination rates. Any scenario where you'd otherwise use Claude Opus 4.8, the quality gap remains significant.
DeepSeek V4's pricing, context length, and MIT license make it a strong foundation for future development. This is a preview release, and the gap between benchmarks and real-world execution may narrow with updates. For now, start with Flash for cost-sensitive work and test Pro only when your own benchmarks show a meaningful quality difference. The models are cheap, which is good. Cheap and reliable would be better, and we're not fully there yet.
