Launch promo: first 30 days free on legacy modelsUnified API + V4 watch

DeepSeek V4 Network

Unified access to the DeepSeek model lineup. Run production workloads on V3.1, R1, Math-7B, Janus-Pro-7B, and VL2 today, and stay ready for V4.

Unified API surface - V4 launch intelligence - transparent pricing

Unified DeepSeek API
Switch between text, reasoning, math, and multimodal models with one unified API surface.
Legacy models, lowest cost
Launch promo includes 30 days free for legacy models, with clear pricing after the trial.
V4 launch tracking
We track community updates, benchmarks, and release signals so you can be ready on day one.

Signal stream

The themes we track daily to anticipate V4 readiness.

Live updates
MoE routing stability
100K context targets
Reasoning traces
Multimodal depth
Inference cost curves
Self-hosting signals
MoE routing stability
100K context targets
Reasoning traces
Multimodal depth
Inference cost curves
Self-hosting signals
MoE routing stability
100K context targets
Reasoning traces
Multimodal depth
Inference cost curves
Self-hosting signals
MoE routing stability
100K context targets
Reasoning traces
Multimodal depth
Inference cost curves
Self-hosting signals
Latency benchmarks
Safety alignment
Tool use reliability
Agentic workflows
Vision + OCR
Video roadmap
Latency benchmarks
Safety alignment
Tool use reliability
Agentic workflows
Vision + OCR
Video roadmap
Latency benchmarks
Safety alignment
Tool use reliability
Agentic workflows
Vision + OCR
Video roadmap
Latency benchmarks
Safety alignment
Tool use reliability
Agentic workflows
Vision + OCR
Video roadmap

Where teams apply DeepSeek V4

High-impact workflows

Use cases where long context, reasoning depth, and multimodal inputs make the biggest impact.

Code generation and refactoring

Generate modules, migrate frameworks, and fix bugs faster with structured reasoning.

Math and scientific reasoning

Solve GSM8K-style problems, technical derivations, and verification tasks.

Multimodal analysis

Combine text and image understanding for documents, charts, and OCR-heavy workflows.

Knowledge-base QA

Query large internal docs with long-context retrieval and structured answers.

Research synthesis

Summarize papers, compare methods, and extract evidence with citations.

Enterprise copilots

Deploy assistant workflows with guardrails, usage caps, and cost controls.

Research brief

DeepSeek V4: architecture and launch signals

DeepSeek V4 is widely discussed as a next-generation Mixture-of-Experts system. Our research brief aggregates public analysis describing a trillion-scale parameter budget with sparse activation per token, a shared expert plus routed experts, and top-k routing to keep inference practical. Those notes also highlight long-context ambitions (100K-class windows are frequently mentioned) for large documents, codebases, and multi-stage reasoning.

Training discussions emphasize bigger and cleaner corpora, heavier math and code weighting, and improved routing balance. Reported benchmark references include MMLU, HumanEval, GSM8K, and MATH, but we treat these figures as directional until verified by official releases or third-party evaluations. The same notes point to multimodal expansion: strong image understanding today and video generation on the roadmap.

While V4 remains pre-launch, the current DeepSeek lineup is production-ready. V3.1 handles general chat and long context, R1 focuses on structured reasoning, Math-7B offers cost-efficient numerical accuracy, Janus-Pro-7B targets multimodal generation, and VL2 excels at OCR and document analysis. We keep these models accessible through a unified API surface and prepare the waitlist for V4 access.

Community watchlist
The most discussed signals we track before launch.
  • Release timing and whether the schedule shifts again.
  • Multimodal depth: image quality now, video generation later.
  • Benchmark verification from official or third parties.
  • Open-source and self-hosting expectations plus chip optimization signals.
  • V4 access model, pricing mechanics, and rollout pace.
Read the latest updates

At-a-glance V4 signals

Reported scale and context

Key reported signals we track while waiting for official specifications.

1T

Total parameters (reported)

320B

Active parameters per token (reported)

100K

Context window (reported)

Model lineup

Production-ready models today, with V4 on deck.

View all models
MultimodalComing soon
DeepSeek V4
Next-generation multimodal MoE model. Launch details and pricing are coming soon.
PricingComing soon
View details
Text
DeepSeek V3.1
Fast, general-purpose MoE model with long-context variants and strong coding ability.
Pricing$1.00 / 1M tokens
View details
Reasoning
DeepSeek R1
Reasoning-first MoE model optimized for multi-step logic, math, and complex planning.
Pricing$1.50 / 1M tokens
View details
Math
DeepSeek Math-7B
Compact math specialist tuned for high-accuracy numeric reasoning and proofs.
Pricing$1.00 / 1M tokens
View details
Multimodal
DeepSeek Janus-Pro-7B
Unified multimodal model for image understanding and text-to-image generation.
Pricing$0.02 / image
View details
Multimodal
DeepSeek VL2
Vision-language model for OCR, documents, charts, and visual Q&A.
Pricing$0.02 / image
View details

Signals from builders

What teams are watching

Jonathan Yombo

ML Engineer

The MoE breakdown makes the trillion-scale claims feel practical for inference.

Yves Kalume

Product Lead

Having V3.1 and R1 behind one endpoint lets us ship now and upgrade later.

Yucel Faruksahan

Researcher

Long-context signals are exactly what we need for paper and dataset synthesis.

Anonymous author

Full-stack Developer

Benchmark tracking with sources keeps the hype in check.

Shekinah Tshiokufila

AI Engineer

The multimodal roadmap is clear: images today, video next.

Oketa Fred

Data Scientist

Math and code performance are front-and-center instead of buried in marketing.

Zeki

Infra Lead

I appreciate the self-hosting and domestic-chip discussion - it matters for deployment.

Joseph Kitheka

Startup Founder

The waitlist flow is simple and keeps our team informed.

Khatab Wedaa

Solutions Architect

Unified API plus usage caps make budgeting predictable.

Rodrigo Aguilar

Developer Advocate

The docs focus on real workflows - code, math, and knowledge bases.

Eric Ampire

Research Engineer

Reported MMLU and HumanEval gains align with what we see internally.

Roland Tubonge

CTO

It is the best balance of community signal and verified data I have seen.

Built for developers and teams

Ship faster with clean docs, transparent pricing, and production-ready tooling for the DeepSeek ecosystem. The goal is straightforward: keep legacy models affordable, expose the same API surface across text, reasoning, math, and vision, and make the V4 rollout predictable through a public waitlist and update cadence.

The product direction mirrors the PRD focus: a single unified endpoint, a clear trial period for legacy models, and a V4 access path that scales from early credits to full launch pricing. If you are evaluating for teams, the Playground is the fastest way to compare outputs, then lock in a plan once you see which model behaves best for your workload.

Unified API surface
Drop-in requests with familiar endpoints and toolchains.
Playground with model switcher
Compare V3.1, R1, Math-7B, Janus-Pro-7B, and VL2 instantly.

Simple, transparent access

Legacy models start with a 30-day free window, then 9.9 / 29.9 / 59.9 monthly tiers. V4 pricing is announced at launch.

Free
$0
Launch promo for legacy models

  • 30 days free access to legacy models
  • Unified API surface
  • Usage caps and rate limits apply
  • Community support
  • No SLA
  • PricePlans.free.limits.limit-2
  • PricePlans.free.limits.limit-3
Lite
$9.9/month
For solo builders and prototypes

  • 10M tokens per month
  • 200 images included
  • 2 requests per second
  • 1 API key
  • Email support
  • Overages billed pay-as-you-go
Popular
Plus
$29.9/month
Best for growing teams

  • 40M tokens per month
  • 800 images included
  • 5 requests per second
  • 3 team seats
  • Priority support
  • Overages billed pay-as-you-go
Pro
$59.9/month
For production workloads

  • 120M tokens per month
  • 2,500 images included
  • 8 requests per second
  • 5 team seats
  • Priority support
  • Overages billed pay-as-you-go

FAQ

Release, access, and benchmarks

Be ready for V4 without waiting

Join the waitlist for launch access, then use V3.1, R1, VL2, Janus-Pro-7B, and Math-7B today with the same API surface.