PricingDescribed by vendor as highly cost-effective API pricing; emphasizes smaller-parameter variant for faster responses and economical usage.
DEEPSEEK · MODEL RELEASE TRACKER
DeepSeek-V4
DeepSeek-V4 is a v4 family release by DeepSeek. The preview was published and open-sourced; it ships in two variants (V4‑Pro and V4‑Flash), supports a 1M-token context as default, and is claimed to offer strong reasoning, world knowledge, and cost-effective long-context efficiency. It was used as a backbone model in SAGE experiments that report large reductions in regression rates on some benchmarks.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Context windowDefault/standard context length is 1,000,000 tokens (1M) across official DeepSeek services; both V4‑Pro and V4‑Flash support 1M.
Not established from the available sources.
Empirical results in SAGE experimentsIn the SAGE paper's experiments, applying SAGE to DeepSeek‑V4 lowered regression rates (example: LiveMath from 36.5% to 0%; OfficeQA from 42.8% to 0%) and increased final scores (example: LiveMath from 34.15 to 48.78); SAGE attained the highest final score in all 20 reported settings.
AvailabilityAPI updated and available; interactive access via chat.deepseek.com (Expert/Instant modes). Supports OpenAI ChatCompletions and Anthropic APIs.
VERIFIABLE FACTS
Every value stays attached to a source and date.
RELEASE · Release statusDEVELOPER CLAIM
Preview announced and open-sourced (DeepSeek announcement).
AVAILABILITY · AvailabilityDEVELOPER CLAIM
API updated and available; interactive access via chat.deepseek.com (Expert/Instant modes). Supports OpenAI ChatCompletions and Anthropic APIs.
CONTEXT WINDOW · Context windowDEVELOPER CLAIM
Default/standard context length is 1,000,000 tokens (1M) across official DeepSeek services; both V4‑Pro and V4‑Flash support 1M.
CAPABILITIES · Model configurations / parameter countsDEVELOPER CLAIM
DeepSeek‑V4‑Pro: 1.6T total / 49B active parameters. DeepSeek‑V4‑Flash: 284B total / 13B active parameters.
CAPABILITIES · Architecture and long-context efficiencyDEVELOPER CLAIM
Introduces token-wise compression and DSA (DeepSeek Sparse Attention); claims world-leading long-context efficiency with reduced compute and memory costs.
CAPABILITIES · Benchmarks / capability claimsDEVELOPER CLAIM
Vendor claims: leads current open models on world knowledge (trailing only Gemini‑3.1‑Pro); beats current open models in Math/STEM/Coding and rivals top closed‑source models on reasoning; smaller variant reportedly performs on par with V4‑Pro on simple agent tasks.
PRICING · PricingDEVELOPER CLAIM
Described by vendor as highly cost-effective API pricing; emphasizes smaller-parameter variant for faster responses and economical usage.
BENCHMARKS · Empirical results in SAGE experimentsINDEPENDENTLY SUPPORTED
In the SAGE paper's experiments, applying SAGE to DeepSeek‑V4 lowered regression rates (example: LiveMath from 36.5% to 0%; OfficeQA from 42.8% to 0%) and increased final scores (example: LiveMath from 34.15 to 48.78); SAGE attained the highest final score in all 20 reported settings.
WHAT CHANGED