Table of Contents
The world of artificial intelligence is on the verge of a massive shift. DeepSeek V4 is the flagship model family from DeepSeek that pushes the boundaries of parameter scale while promising unprecedented efficiency. V4-Pro totals 1.6 trillion parameters (~49B active per token via MoE), while V4-Flash totals 284 billion parameters (~13B active). Both are text-only and offer a 1 million token context window, positioning them as direct competitors to Western giants like OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.8.
In this comprehensive article, we dive into the key specifications, architectural innovations, expected pricing, and the strategic hardware shifts behind DeepSeek V4.
Key Takeaways
- •V4-Pro 1.6T / V4-Flash 284B parameters with ~49B / ~13B active per token via MoE
- •1 Million token context window — equivalent to 15–20 full novels
- •Text-only: no vision, audio or video modality in either shipped variant
- •10–50x cheaper API pricing than GPT-5.4 and Claude Opus 4.8
- •Open-source weights released under the MIT license
- •Runnable locally on dual RTX 4090s or single RTX 5090
The DeepSeek V4 Family in 2026
DeepSeek V4 launched on 24 April 2026 as a two-model family. There is no separate "base" model alongside them — the official release and API docs list exactly two variants: V4 Pro (1.6T total / 49B active) and V4 Flash (284B total / 13B active). Both share the same architecture, 1M-token context, and MIT-licensed open weights.
DeepSeek V4 Pro
1.6T total / 49B active MoE. Production-tuned tier. As of 31 May 2026, the previous 75% price cut became permanent: $0.435/M input (cache miss), $0.87/M output.
Read the price-cut analysis →DeepSeek V4 Flash
284B total / 13B active MoE. Latency-optimised variant with the same 1M context. $0.14/M input (cache miss), $0.28/M output.
V4 Flash deep dive →Both variants share the same V4 architecture covered below — MoE with 13B (Flash) or 49B (Pro) active params per token, text-only inputs, and DeepSeek Sparse Attention. The differences are model size, latency, and pricing.
What is DeepSeek V4? (Core Specs)
DeepSeek V4 builds upon the success of its predecessors (like V3 and R1) by combining massive scalability with extreme operational cost efficiency. Here are the primary technical specifications:
Parameters
V4-Pro 1.6T total / 49B active · V4-Flash 284B / 13B (MoE)
Active Parameters
13B (Flash) or 49B (Pro) activated per token via efficient Mixture-of-Experts (MoE) routing
Context Window
1 Million tokens — roughly an entire medium-sized codebase or 15–20 full-length novels
Text-only
Both V4 variants accept text input only — no vision, audio or video modality
Architecture: what V4 actually changes
V4's efficiency comes from two things DeepSeek documents in the model cards — a stability framework for training a 1.6T-parameter MoE model, and a sparse attention stack for long context. A third line of work, Engram, is frequently attached to V4 online; it is a separate research paper, not a shipped V4 component.
1. Manifold-Constrained Hyper-Connections (mHC)
As AI models scale up, they often suffer from training instability (such as gradient explosion). mHC is a mathematical framework that constrains signal amplification, keeping it under 2x (compared to an unconstrained 3000x).
⚡ This allows DeepSeek to stably train a 1.6T-parameter MoE model with only a 6.7% computational overhead.
2. DeepSeek Sparse Attention (DSA)
To process 1 million tokens affordably, V4 replaces dense attention with DeepSeek Sparse Attention. The V4 model cards describe it in terms of compressed sparse attention (CSA) and hierarchical/coarse-grained attention (HCA): the context is compressed into coarse summaries, the relevant regions are selected, and full attention is applied only there. The "Lightning Indexer" often quoted alongside DSA belongs to V3.2-Exp, not V4.
🚀 The result is roughly linear rather than quadratic cost growth in long-context work, which is what makes 1M-token pricing viable.
Related research: Engram conditional memory (not a V4 feature)
Engram is a DeepSeek research paper on decoupling static factual recall from dynamic reasoning (arXiv:2601.07372). It is not part of the shipped V4-Pro or V4-Flash architecture, and DeepSeek has not stated that it will be. Treat reported Engram retrieval gains as results from that paper, not as V4 specifications.
Benchmarks: DeepSeek V4 vs. GPT-5.4 and Claude 4.5
DeepSeek V4 is heavily focused on software engineering and deep reasoning. According to leaked internal benchmarks, the model performs at an extraordinary level:
| Benchmark | DeepSeek V4 | GPT-5.4 | Claude Opus 4.8 | DeepSeek V3 |
|---|---|---|---|---|
| SWE-bench Verified | >80% | ~80% | 88.6% | ~49% |
| Context Window | 1M tokens | ~1M | 1M | 128K |
| Parameters (Total / Active) | 1.6T / 49B (Pro) | Undisclosed | Undisclosed | 671B / 37B |
⚠️ Note: These impressive numbers are currently based on leaked internal data and are awaiting independent third-party verification upon release.
API Pricing: The Most Cost-Effective Frontier AI
Western frontier models are powerful but expensive. GPT-5.4 costs $2.50/M input ($0.25/M cached) and $15.00/M output under 270K context; Claude Opus 4.8 costs $5.00/M input and $25.00/M output. DeepSeek V4's published rates sit an order of magnitude below both:
| Pricing Tier (per 1M tokens) | DeepSeek V4-Flash | DeepSeek V4-Pro | GPT-5.4 | Claude Opus 4.8 |
|---|---|---|---|---|
| Input (cache miss) | $0.14 | $0.435 | $2.50 | $5.00 |
| Input (cache hit) | $0.0028 | $0.003625 | $0.25 | $2.50 |
| Output | $0.28 | $0.87 | $15.00 | $25.00 |
DeepSeek rates verified 2026-07-25; competitor rates verified 2026-07-25. GPT-5.4 rates apply under 270K context. See the full DeepSeek pricing breakdown.
💡 Measured against V4-Pro at $0.435/$0.87 per 1M tokens, the actual multiples are:
- • vs GPT-5.4: 5.7× on input, 17.2× on output
- • vs Claude Opus 4.8: 11.5× on input, 28.7× on output
Geopolitics and Hardware: The Shift to Huawei
One of the most consequential aspects of DeepSeek V4 is its hardware foundation. Due to strict US export restrictions on advanced Nvidia GPUs (like the B300 and H200), DeepSeek has optimized V4 to run heavily on domestic Chinese silicon for inference.
While initial training likely still utilized Nvidia hardware (such as H800s), the model is highly optimized for the Huawei Ascend 950PR and Cambricon MLU chips.
🔧 Huawei's Ascend 950PR reportedly delivers 2.87x the compute performance of the Nvidia H20 (the chip legally allowed for export to China). This marks a major milestone in China's push for AI semiconductor independence.
DeepSeek V4 Release Timeline: From Launch to V4 Pro
DeepSeek V4 launched on 24 April 2026 as a two-model family (Pro and Flash). Below is the timeline from launch through the V4 Pro pricing decision and the retirement of the legacy model IDs.
🚀 April 24, 2026 — V4 Pro & V4 Flash launch: DeepSeek shipped exactly two V4 variants: V4 Pro (1.6T total / 49B active) and V4 Flash (284B / 13B), both text-only, both with a 1M-token context window and MIT-licensed open weights.
💸 May 31, 2026 — V4 Pro price cut made permanent: The 75% discount on V4 Pro became permanent at $0.435/M input (cache miss) and $0.87/M output. Read the full analysis.
📅 July 24, 2026 — Legacy model IDs retired: deepseek-chat and deepseek-reasoner were retired at 15:59 UTC. They are no longer routed anywhere — see what the retirement did to your bill.
MoE routing cuts compute per token, not memory: every expert still has to be resident. V4 Flash's 284B total parameters are roughly 140 GB+ of weights in the shipped mixed precision, and still 70–80 GB heavily quantised — so local inference means multi-GPU server hardware, not a consumer desktop. V4 Pro at 1.6T total params is firmly a data-centre workload. For most teams the hosted API is cheaper than self-hosting once you account for GPU capital, electricity and ops time.
Frequently Asked Questions About DeepSeek V4
Conclusion
Three months after launch, DeepSeek V4 has done exactly what the leaks predicted: architectural efficiency (mHC plus DeepSeek Sparse Attention with CSA/HCA) beating brute-force scaling on cost-per-quality. With a 1-million-token context window, the V4 Pro permanent price cut, and MIT-licensed open weights, V4 has forced every major API provider to revisit pricing. Both shipped variants are text-only. For most teams in mid-2026 the practical choice is V4 Pro for the hardest reasoning and V4 Flash for low-latency, high-throughput agentic use — the two variants DeepSeek actually shipped.