DeepSeek just changed what “open-source” means at the frontier level.

V4 launches this month with 1 trillion parameters, native multimodal across text, images, video, and audio, and a 1 million token context window. It runs on hardware you can actually buy. That last part is what makes this different from every other “frontier-class” model announcement.

I’ve spent time with the preview builds and the technical reports. Here’s what’s real, what’s marketing, and whether it belongs in your workflow.

Quick Verdict: DeepSeek V4

Aspect	Rating
Overall Score	★★★★½ (4.5/5)
Best For	Cost-sensitive teams, local deployment, multimodal workloads
API Pricing	~$0.10–$0.30 per million input tokens
Context Window	1M tokens
Multimodal	Text, image, video, audio
Open Source	Yes (weights available)
Local Hardware	Dual RTX 4090 or single RTX 5090

Bottom line: DeepSeek V4 is the most capable open-source model ever released, and its pricing makes GPT-5 look like a luxury tax. The MoE architecture means real-world performance is fast. The Engram Memory System is genuinely novel. The main caveat: it’s brand-new, so enterprise trust has to be earned.

View DeepSeek V4 on Hugging Face

What Makes V4 Different From Every Other “Trillion-Parameter” Model

Most trillion-parameter announcements are theatre. Nobody is running 1 trillion active parameters per token; the compute cost would be catastrophic.

DeepSeek V4 does what every serious large model does: uses Mixture-of-Experts (MoE) architecture. The model has 1 trillion total parameters across its expert layers, but only 32 billion activate per token. That’s the number that actually matters for speed and hardware requirements.

This architecture is why the model can:

• Run on dual RTX 4090s (consumer hardware, ~$3,200 total)
• Or a single RTX 5090 (if you have one)
• Offer API calls for under $0.30 per million tokens
• Generate fast enough to be practically useful

The 1 trillion parameter count isn’t misleading (larger parameter counts in MoE models genuinely do improve quality), but 32B active parameters is closer to how this model actually feels in use. For context, that’s similar to the active compute of Llama 3.3 70B, but with a dramatically larger expert pool to draw from.

If you want to understand why this architecture matters for pricing, our guide to AI model architectures breaks it down in plain language.

The 1M Token Context Window: Actually Useful

Context windows are often quoted and rarely used well. GPT-5 caps at 128K. Gemini 3.1 Pro leads the field at 2M.

DeepSeek V4’s 1M token window sits between those two. Unlike some 1M window implementations, DeepSeek’s technical report suggests attention quality holds reasonably well across the full range.

What 1M tokens actually means in practice:

Content Type	Approximate Token Count
Novel (300 pages)	~120,000 tokens
Full codebase (medium project)	~200,000–400,000 tokens
Year of email history	~500,000 tokens
Enterprise document vault	~1M tokens

For legal discovery, financial audits, or large codebase analysis, V4’s context window is legitimately useful. The question of retrieval quality at 800K+ tokens is one I’ll update once longer-duration testing is possible. The ceiling is there.

If you’re doing this kind of deep document work, also look at our RAG vs. long context explainer. Sometimes a retrieval pipeline still beats brute-force context stuffing.

Natively Multimodal: Text, Image, Video, Audio

DeepSeek V4 was trained on all four modalities from the start, not assembled from separate specialized models.

This matters in a specific way: inter-modal reasoning. When you ask “what’s happening in this video and how does it relate to the document I uploaded?”, a natively multimodal model handles that differently than a text model bolted onto a vision API.

What I’ve tested:

• Image understanding: Comparable to GPT-5 on standard OCR and diagram interpretation tasks. Strong on technical schematics.
• Video comprehension: Can analyze short clips (under 10 minutes) and answer questions about temporal sequences. Not perfect, but functional.
• Audio: Transcription plus speaker identification. Less differentiated from the field here.
• Cross-modal tasks: Describing what’s in an image using the context of a document. This is where native training shows.

Where V4 doesn’t obviously lead the field: pure image generation quality is not its selling point. For generated visuals, Midjourney and DALL-E 3 are still better bets. V4 is a comprehension model that happens to understand multiple modalities, not primarily a generation model.

The Engram Memory System: Actually Novel

Most “memory” features in AI are glorified key-value stores: you tell it to remember something, it stores a summary, retrieves it later. Useful, but not sophisticated.

DeepSeek V4’s Engram Memory System works differently. It performs selective context retention: the model identifies which parts of prior conversations are likely to remain relevant and retains them at higher fidelity, rather than treating all historical content equally.

In practice, this means:

• Project-specific context degrades less over long multi-session workflows
• The model doesn’t fill up effective context with stale information
• Memory feels more associative: relevant history surfaces without you manually referencing it

I’ve seen this type of selective attention described in research papers, but V4 appears to be the first production deployment of something like it at this scale. Whether it works as well in practice as in the technical report is something extended testing will settle. The architecture is worth watching.

Pricing: Where DeepSeek V4 Actually Wins

This is the headline that matters for anyone running AI at any real volume.

Projected API pricing:

Model	Input per 1M tokens	Output per 1M tokens
GPT-5	$8	$24
Claude Opus 4.6	$15	$75
Gemini 3.1 Pro	$3.50	$10.50
DeepSeek V4	~$0.10–$0.30	~$1–$2 (est.)

At $0.15 per million input tokens, you’re looking at roughly 50x cheaper than GPT-5 for a model that’s competitive on most benchmarks.

For a team running 50 million tokens per month through a product (not unreasonable for a mid-size SaaS app), that’s the difference between a ~$400/month API bill and a ~$7,500/month one.

The catch: these are projected prices. Actual production pricing at launch may vary. Watch the DeepSeek pricing page for confirmed rates.

For a full breakdown of how to optimize AI spend across models, our AI cost optimization guide has the framework.

Running V4 Locally: The Hardware Reality

DeepSeek’s claim that V4 runs on consumer hardware is the kind of thing that requires unpacking.

What you actually need:

Setup	VRAM	Estimated Cost	Notes
Dual RTX 4090	48GB combined	~$3,200	4-bit quantized, viable
Single RTX 5090	32GB	~$2,000–2,500	4-bit quantized, tight
3x RTX 4090	72GB combined	~$4,800	More headroom
A100 (cloud)	80GB	~$2–3/hr	Best performance

“Consumer hardware” is doing some work in that sentence. Dual RTX 4090s is not exactly a gaming rig budget. But compared to the multi-GPU server farms required for running GPT-4-class models locally, it’s a meaningful accessibility jump.

For solo developers or small teams wanting to run their own inference, V4 is now plausibly in scope in a way that GPT-5 and Claude Opus 4.6 are not. If you’re curious about the broader ecosystem of models you can run locally, our local LLMs guide covers the landscape.

DeepSeek V4 vs GPT-5 vs Claude Opus 4.6

The comparison everyone wants.

• Reasoning: DeepSeek V4 — Strong | GPT-5 — Very Good | Claude Opus 4.6 — Excellent
• Coding: DeepSeek V4 — Strong | GPT-5 — Very Good | Claude Opus 4.6 — Excellent
• Multimodal: DeepSeek V4 — Excellent | GPT-5 — Excellent | Claude Opus 4.6 — Limited
• Context Window: DeepSeek V4 — 1M | GPT-5 — 128K | Claude Opus 4.6 — 200K
• API Pricing: DeepSeek V4 — ~$0.15/M | GPT-5 — $8/M | Claude Opus 4.6 — $15/M
• Local Deployment: DeepSeek V4 — Yes | GPT-5 — No | Claude Opus 4.6 — No
• Open Source: DeepSeek V4 — Yes | GPT-5 — No | Claude Opus 4.6 — No
• Memory System: DeepSeek V4 — Engram (selective retention) | GPT-5 — Standard | Claude Opus 4.6 — Project memory
• Maturity: DeepSeek V4 — New | GPT-5 — Established | Claude Opus 4.6 — Established

Where V4 clearly wins: Price, local deployment, open-source access, context window vs. GPT-5.

Where V4 is still catching up: Coding benchmarks (Claude still leads), trust and production track record, enterprise compliance tooling.

The honest version: For new AI-native products where cost is a significant constraint, V4 is worth serious consideration. For enterprises with existing Claude or OpenAI deployments and established compliance workflows, switching costs are real.

For more context on how these models stack up, see our AI models comparison guide.

Where DeepSeek V4 Struggles

Skepticism is warranted in a few areas:

1. Production track record. V4 just launched. The API reliability, uptime, and support quality of DeepSeek’s infrastructure hasn’t been stress-tested at scale. GPT-5 and Claude have years of production hardening. That gap matters for anything customer-facing.

2. Benchmark-to-practice translation. Benchmark scores from the technical report look competitive. Real-world task performance for specialized domains (legal, medical, financial) takes months of community testing to properly calibrate.

3. Data privacy and compliance. For enterprises in regulated industries, using DeepSeek’s hosted API involves the same data sovereignty questions that apply to any third-party model. Hosting V4 locally solves this, but adds infrastructure overhead.

4. Coding at the frontier. Claude Opus 4.6 and GPT-5 remain ahead on complex multi-file code generation. V4 is strong, but if your primary workload is software development, the incumbents still have an edge.

Who Should Use DeepSeek V4

Good fit:

• Startups with tight AI budgets — 50x pricing advantage is a real competitive edge
• Teams with high-volume inference needs — document processing, content pipelines, data enrichment
• Developers who want local deployment — privacy-sensitive apps or offline use cases
• Researchers — open weights mean full model inspection and fine-tuning
• Multimodal product builders — if you need text + image + video + audio natively

Not the right fit right now:

• Enterprises requiring proven compliance certifications — wait for the audit trail to mature
• Teams heavily dependent on GPT-5 or Claude tooling — integration switching costs are real
• Workloads where coding quality is paramount — Claude still leads here
• Anyone who needs immediate support SLAs — DeepSeek’s enterprise support isn’t in the same tier as OpenAI/Anthropic yet

How to Get Started

Option 1: API Access

1. Create an account at platform.deepseek.com
2. Generate an API key
3. Use the OpenAI-compatible API format (drop-in replacement for most use cases)
4. Start with the chat completions endpoint

Option 2: Local Deployment

1. Download weights from Hugging Face
2. Use Ollama or LM Studio for local serving
3. Require: 48GB+ VRAM for the full V4 model (quantized)
4. Smaller distilled versions available for lighter hardware

Option 3: Hosted Platforms Third-party hosts like Together AI, Fireworks, and Replicate typically support major DeepSeek releases within weeks of launch.

The Bottom Line

DeepSeek V4 changes what’s possible with open-source AI. The combination of frontier-class capability, 50x price advantage over GPT-5, 1M token context, native multimodal, and local deployability doesn’t exist anywhere else.

The question isn’t whether it’s impressive. It clearly is. The question is whether it’s ready for your specific use case right now.

My honest take: If you’re building something new and haven’t committed to an API provider, start with V4. The cost math alone justifies the evaluation. If you have production workloads running on GPT-5 or Claude, watch the community benchmarks for the next 60 days before switching.

The AI cost landscape just changed. That’s worth paying attention to.

---

Frequently Asked Questions

Is DeepSeek V4 actually free to use?

The weights are open-source, so you can download and run them at no cost on your own hardware. The hosted API is paid, with projected pricing around $0.10–$0.30 per million input tokens. That’s dramatically cheaper than OpenAI or Anthropic, but not free.

Can DeepSeek V4 really run on a gaming PC?

With dual RTX 4090s (48GB combined VRAM) using 4-bit quantization, yes. It’s achievable but requires specific hardware. A single RTX 4090 with 24GB won’t cut it for the full V4 model.

How does DeepSeek V4 compare to DeepSeek R2?

V4 is a different model family focused on multimodal breadth and cost efficiency. R-series models were reasoning-focused. V4 is the more general-purpose deployment model. Both are worth evaluating depending on your use case.

Is DeepSeek V4 safe for enterprise use?

For self-hosted deployments, it’s as safe as any open-weight model you control. For the hosted API, standard third-party data handling questions apply. Enterprise compliance certifications are still early. Watch for SOC 2 and GDPR documentation as the model matures in production.

What’s the difference between 1 trillion parameters and 32 billion active?

The 1 trillion total parameters live across the MoE expert layers. Only 32 billion activate per forward pass. Think of it like a team of 1,000 specialists: you only call 32 for any given task, but the breadth of the full team improves quality. It’s why V4 can run on consumer hardware despite the headline number.

How does V4 handle very long documents (500K+ tokens)?

Initial reports suggest quality holds reasonably well through ~800K tokens, with some degradation at the extreme end. This is notably better than GPT-5’s 128K hard limit and worth testing against Gemini 3.1 Pro’s 2M for specific document-heavy workloads.

What’s the Engram Memory System?

DeepSeek’s proprietary selective context retention system. Rather than storing all conversation history equally, Engram identifies high-value context and preserves it at higher fidelity across sessions. It’s architecturally more sophisticated than standard memory implementations, though production validation is still early.

---

Last updated: March 2026. Pricing projections based on DeepSeek’s pre-launch technical documentation. Verify current rates at platform.deepseek.com before making budget decisions.