Compare the listed model data for DeepSeek-V3 by DeepSeek and DeepSeek-V3.1 by DeepSeek. Benchmark and pricing records are not a controlled test and may change.
Data last updated: 2026-10-16. Verify current model details with each provider.
What the listed data says
DeepSeek-V3: 128K tokens; DeepSeek-V3.1: 128K tokens. The listed context-window values are equal. DeepSeek-V3 is listed for coding, agentic workflows, long-context tasks; DeepSeek-V3.1 is listed for no specific capability flags. Shared catalog benchmark entries: MMLU: DeepSeek-V3 88.5%; DeepSeek-V3.1 91.7%; HumanEval: DeepSeek-V3 89.9%; DeepSeek-V3.1 93.8%; LiveCodeBench: DeepSeek-V3 51.4%; DeepSeek-V3.1 68.2%; SWE-bench Verified: DeepSeek-V3 42.0%; DeepSeek-V3.1 52.3%. These records are reference data, not a controlled head-to-head test; verify current pricing and specifications with each provider.
Catalog source links
Use these model and provider links to verify current specifications, pricing, and availability.
The listed context windows are 128K tokens and 128K tokens; the larger is 1× the smaller.
On shared numeric benchmark entries, MMLU: DeepSeek-V3 − DeepSeek-V3.1 = -3.2 percentage points (88.5% vs 91.7%); HumanEval: DeepSeek-V3 − DeepSeek-V3.1 = -3.9 percentage points (89.9% vs 93.8%); LiveCodeBench: DeepSeek-V3 − DeepSeek-V3.1 = -16.8 percentage points (51.4% vs 68.2%); SWE-bench Verified: DeepSeek-V3 − DeepSeek-V3.1 = -10.3 percentage points (42.0% vs 52.3%).
The first listed parameter quantities are 671B total (37B active via MoE) and 671B total (37B active via MoE); the larger quantity is 1× the smaller. Total and active MoE parameters are not directly equivalent.
The catalog release dates are December 2024 and June 2025, about 6 months apart at the precision supplied.
Cost example: 1M input and 1M output tokens
A cost example for 1M input plus 1M output tokens cannot be calculated reliably because separate numeric input and output rates are not listed for both models.
Choose DeepSeek-V3 if…
Choose DeepSeek-V3 when your task aligns with catalog-listed uses or strengths such as General-purpose coding and software development, Business document analysis, API building at high quality and low cost, Enterprise AI integration at scale.
Choose DeepSeek-V3.1 if…
Choose DeepSeek-V3.1 when your task aligns with catalog-listed uses or strengths such as Enterprise software development requiring high quality at low latency, Document intelligence and knowledge base applications, API providers serving general business customers, DeepSeek's most capable non-reasoning model as of June 2025.
Catalog specifications for DeepSeek-V3 and DeepSeek-V3.1
Specification
DeepSeek-V3
DeepSeek-V3.1
Provider
DeepSeek
DeepSeek
Parameters
671B total (37B active via MoE)
671B total (37B active via MoE)
Context window
128K tokens
128K tokens
Maximum output
Not listed
Not listed
Catalog pricing
MIT open source. DeepSeek API: $0.27/MTok input, $1.10/MTok output
MIT open source. DeepSeek API: ~$0.27/MTok input, ~$1.10/MTok output
Listed input/output cost per 1M tokens
input $0.27, output $1.1 per 1M tokens
Separate numeric input/output rates not listed
License
MIT
MIT
Coding support
Listed
Not listed
Reasoning support
Not listed
Not listed
Vision input
Not listed
Not listed
Multimodal
Not listed
Not listed
Open weights
Yes
Yes
Benchmark: MMLU
88.5%
91.7%
Benchmark: HumanEval
89.9%
93.8%
Benchmark: LiveCodeBench
51.4%
68.2%
Benchmark: SWE-bench Verified
42.0%
52.3%
Figures are transcribed from LITHOVEX model records. Benchmark versions and evaluation setups may differ; pricing and specifications can change. Check each provider's current documentation before making a decision.
Frequently Asked Questions
What is different between DeepSeek-V3 and DeepSeek-V3.1?
DeepSeek-V3: 128K tokens; DeepSeek-V3.1: 128K tokens. The listed context-window values are equal. DeepSeek-V3 is listed for coding, agentic workflows, long-context tasks; DeepSeek-V3.1 is listed for no specific capability flags. Shared catalog benchmark entries: MMLU: DeepSeek-V3 88.5%; DeepSeek-V3.1 91.7%; HumanEval: DeepSeek-V3 89.9%; DeepSeek-V3.1 93.8%; LiveCodeBench: DeepSeek-V3 51.4%; DeepSeek-V3.1 68.2%; SWE-bench Verified: DeepSeek-V3 42.0%; DeepSeek-V3.1 52.3%. These records are reference data, not a controlled head-to-head test; verify current pricing and specifications with each provider.
What benchmark values are listed for DeepSeek-V3 and DeepSeek-V3.1?
MMLU: DeepSeek-V3 88.5%; DeepSeek-V3.1 91.7%; HumanEval: DeepSeek-V3 89.9%; DeepSeek-V3.1 93.8%; LiveCodeBench: DeepSeek-V3 51.4%; DeepSeek-V3.1 68.2%; SWE-bench Verified: DeepSeek-V3 42.0%; DeepSeek-V3.1 52.3%. The catalog values may use different versions and evaluation setups, so they are reference points rather than a controlled head-to-head result.
Which model has the larger listed context window?
DeepSeek-V3: 128K tokens; DeepSeek-V3.1: 128K tokens. The listed context-window values are equal.
What are the listed costs per million tokens for DeepSeek-V3 and DeepSeek-V3.1?
DeepSeek-V3: input $0.27, output $1.1 per 1M tokens (MIT open source. DeepSeek API: $0.27/MTok input, $1.10/MTok output). DeepSeek-V3.1: Separate numeric input/output rates not listed (MIT open source. DeepSeek API: ~$0.27/MTok input, ~$1.10/MTok output). Where separate numeric input and output rates are available, the worked example uses 1M input plus 1M output tokens. Confirm current terms with each provider.
How far apart are the listed release dates for DeepSeek-V3 and DeepSeek-V3.1?
The catalog lists DeepSeek-V3 as December 2024 and DeepSeek-V3.1 as June 2025, about 6 months apart at the precision supplied.
When was the comparison data last updated?
The model catalog records 2026-10-16 as its data update date. Provider specifications and prices may have changed since then.
How should I choose between DeepSeek-V3 and DeepSeek-V3.1?
Compare the listed specifications and capabilities, review the benchmark notes, then test both with the same representative prompt for your task. This page does not claim a universal winner.