DeepSeek V3
by DeepSeek
The 671B open-weights MoE model that reset the industry's assumptions about training cost
DeepSeek V3 is an open-weights Mixture-of-Experts language model with 671 billion total parameters, of which 37 billion activate per token. Released in December 2024 under a commercially permissive licence, it matched or beat GPT-4o and Claude 3.5 Sonnet on several reasoning and maths benchmarks while reporting a training run practitioners priced at under $6 million.
DeepSeek V3 is a 671-billion-parameter Mixture-of-Experts language model released as open weights on 26 December 2024, of which only 37 billion parameters activate per token. Architecturally it combines Multi-head Latent Attention with the DeepSeekMoE design, an auxiliary-loss-free load-balancing strategy and a multi-token-prediction training objective; it was pre-trained on 14.8 trillion tokens using 2.788 million H800 GPU hours in total, 2.664 million of them for pre-training, in a run its authors report completed without a single irrecoverable loss spike or rollback. Its significance was economic as much as technical: a training bill enterprise architects put at under $6 million produced a model that beat GPT-4o on AIME 2024 (39.2% versus 9.3%) and MATH-500 (90.2 versus 74.6) and edged Claude 3.5 Sonnet on Arena-Hard (85.5 versus 85.2) and AlpacaEval 2.0 (70.0 versus 52.0). Context length is 128K tokens. The code is MIT-licensed and the model licence permits commercial use, and the checkpoint is still pulled more than 1.1 million times a month from Hugging Face against 104,000 GitHub stars. Serving it requires SGLang, vLLM, LMDeploy, TensorRT-LLM or DeepSeek's own inference demo — Hugging Face Transformers never supported it directly — and it runs on NVIDIA, AMD and Huawei Ascend hardware. Two facts govern any 2026 evaluation. First, DeepSeek's own API no longer serves it: the deepseek-chat and deepseek-reasoner aliases were remapped to DeepSeek-V4-Flash when V4 shipped on 24 April 2026 and the legacy names were discontinued on 24 July 2026, so V3 today means self-hosting or a third-party host such as OpenRouter at roughly $0.26 per million input tokens. Second, NIST's CAISI evaluation of October 2025 found DeepSeek models markedly weaker on security and aligned with CCP narratives, which is why many enterprise procurement teams have blocked the name outright regardless of where the weights run.
An infrastructure or ML platform team that already owns GPU capacity and wants a frontier-class open-weights model running entirely inside its own perimeter, with no per-token bill and no data leaving the building.
GPT-4o-class reasoning and maths performance from weights you host yourself under a commercially permissive licence, at hardware cost rather than API cost.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Free, Usage-based
- Target Market
- ML Platform Engineers, CTOs, Data Scientists, Infrastructure Architects, AI Researchers
- Deployment
- Open-source, Self-hosted, API-based
- Founded
- 2023
- Headquarters
- China
Key Features
- ✓671B Mixture-of-Experts with 37B active parameters
Frontier-scale capacity at a fraction of the per-token compute, because only a small expert subset activates for each token.
- ✓Multi-head Latent Attention (MLA)
Compresses the key-value cache so long-context inference consumes far less memory than a standard attention implementation.
- ✓Auxiliary-loss-free load balancing
Balances expert routing without the auxiliary loss that normally degrades MoE model quality, a notable result in the V3 technical report.
- ✓Multi-token prediction objective
Trains the model to predict several tokens ahead, improving data efficiency and enabling speculative decoding at inference time.
- ✓128K-token context window
Handles long documents and large retrieval result sets in a single request without aggressive chunking.
- ✓Permissive licensing
MIT-licensed code plus a model licence that supports commercial use, so the weights can be deployed in production products.
- ✓Broad inference-stack support
Runs under SGLang, vLLM, LMDeploy and TensorRT-LLM on NVIDIA, AMD and Huawei Ascend hardware, with native FP8 weights.
Capabilities
Use Cases
- •On-premises assistant for sensitive data
Run a frontier-class model entirely inside the corporate perimeter so regulated documents and customer records never leave owned infrastructure.
- •High-volume batch inference
Process large document or classification batches on owned GPUs where a per-token API bill would dominate the project's economics.
- •Domain fine-tuning
Adapt an open frontier checkpoint to a specialist corpus, which closed API models either prohibit or price out of reach.
- •Maths and reasoning workloads
Serve quantitative and competition-style reasoning tasks where V3 measured well ahead of GPT-4o on AIME 2024 and MATH-500.
- •Commercial negotiation benchmark
Use measured open-model performance as a price anchor when renegotiating closed-model contracts with incumbent vendors.
Ideal For
Best For
- ✓Self-hosted inference on owned GPU capacity where per-token API pricing is the main cost driver
- ✓Air-gapped or sovereign deployments that cannot send prompts to any hosted model provider
- ✓Maths and reasoning-heavy workloads, where it outscored GPT-4o on AIME 2024 and MATH-500
- ✓Research and fine-tuning on a frontier-scale MoE checkpoint under a licence that permits commercial use
- ✓Cost benchmarking, as a price anchor against incumbent closed-model contracts
Not Ideal For
- ✗Regulated enterprises using DeepSeek's hosted API: prompts are processed in China with no US or EU region option, which conflicts with common data-residency commitments
- ✗Organisations whose procurement has blocked Chinese-origin models outright — practitioners report customers banning 'anything with the word DeepSeek', hosted or self-hosted
- ✗Teams without serious hardware: this is a 671B checkpoint, and practitioners describe needing dual EPYC CPUs with roughly 400GB of RAM or multi-GPU servers for a usable setup
- ✗Anyone wanting a supported vendor API, since DeepSeek retired the V3-series endpoints in July 2026 in favour of V4
- ✗Use cases touching Chinese history, politics or foreign relations, where NIST found censorship consistent with CCP narratives in both English and Chinese
Integrations
Deployment
Market Analysis
Pros
- ✓Beat GPT-4o on AIME 2024 (39.2% vs 9.3%) and MATH-500 (90.2 vs 74.6) and edged Claude 3.5 Sonnet on Arena-Hard and AlpacaEval 2.0
- ✓Open weights under a commercially permissive licence, with MIT-licensed code, so it can ship inside a product
- ✓Dramatically cheaper inference than closed frontier models — one architect measured MMLU parity with o1-mini at 85% lower cost per token
- ✓Still one of the most widely used open checkpoints, with over 1.1 million Hugging Face downloads a month and 104k GitHub stars
- ✓Runs on NVIDIA, AMD and Huawei Ascend hardware with native FP8 weights, so it is not locked to one vendor's silicon
Cons
- ✗DeepSeek's own API no longer serves it: V3-series endpoints were remapped to V4-Flash in April 2026 and the legacy model names were discontinued on 24 July 2026
- ✗NIST's CAISI evaluation (October 2025) found agents built on DeepSeek's most secure model roughly 12 times more likely than US frontier models to follow malicious instructions, with V3.1 complying with phishing prompts 48% of the time versus 0% for the leading US model
- ✗The same evaluation found censorship consistent with CCP narratives across 190 questions on Chinese history, politics and foreign relations, in both English and Chinese
- ✗Hosted use processes data in China with no US or EU region option, which conflicts with common data-residency and compliance commitments
- ✗Self-hosting is genuinely hard: Hugging Face Transformers never supported it directly, and practitioners report needing dual EPYC CPUs with about 400GB of RAM or multi-GPU servers before it is usable
- ✗Procurement bans are common — one practitioner reports customers blocking 'anything with the word DeepSeek', hosted or self-hosted, on origin rather than technical grounds
Pricing
Open weights (self-hosted)
$0
- ✓MIT-licensed code
- ✓Model licence permits commercial use
- ✓FP8 and BF16 weights on Hugging Face
- ✓No usage metering — you pay for hardware
Third-party hosted inference (e.g. OpenRouter)
From $0.26/1M input tokens
- ✓Roughly $0.26 per million input tokens and $1.03 per million output tokens
- ✓Multiple competing providers
- ✓No DeepSeek account required
The weights themselves are free: MIT-licensed code plus a model licence permitting commercial use, so the real cost is GPU capacity rather than tokens, and practitioners describe roughly $10,000 of hardware or dual-EPYC servers with about 400GB of RAM for a workable self-hosted setup. DeepSeek's own API no longer lists V3 — its current catalogue is deepseek-v4-flash at $0.14 per million input tokens (cache miss) and $0.28 output, and deepseek-v4-pro at $0.435/$0.87, with a peak/off-peak policy that doubles prices during Beijing business hours. To buy V3 inference today you go to a third-party host such as OpenRouter, where it runs around $0.26 per million input and $1.03 per million output tokens across competing providers.
Security & Compliance
Connect
Sources
This page was written from 9 sources, 8 on domains other than github.com.
- 1.github.com — DeepSeek V3vendor
- 2.huggingface.co — DeepSeek V3
- 3.huggingface.co — deepseek ai
- 4.api-docs.deepseek.com — updates
- 5.api-docs.deepseek.com — pricing
- 6.openrouter.ai — deepseek chat
- 7.informationweek.com — will enterprises adopt deepseek
- 8.dailysecurityreview.com — nist flags deepseek adoption over security censorship and co
- 9.hn.algolia.com — hn.algolia.com
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Scaled Cognition
APT, a large action model trained to take actions instead of predicting text
Engram
A learned memory layer that makes AI actually know your organization — at up to 100x fewer tokens.
TwelveLabs
Video intelligence API that makes every hour of enterprise footage searchable, analyzable and agent-ready.
Mistral OCR 4
Structure-aware document AI that returns bounding boxes, typed blocks, and per-word confidence scores.