Model Tracker

DeepSeek-V3

DeepSeek-V3 is an AI model made by DeepSeek AI. It is generally available. It takes text and returns text. Its context window is 128,000 tokens.

In inglese

DeepSeek AI

DeepSeek-V3 is an AI model made by DeepSeek AI. It is generally available. It takes text and returns text. Its context window is 128,000 tokens. Its weights are open.

Released: Not confirmed: two readings disagree: 2024-12-26 and 2026-10-16
Status: Generally available — github.com, read 16 October 2026
Takes: Text — github.com, read 16 October 2026
Returns: Text — github.com, read 16 October 2026
Open weights: Yes — github.com, read 16 October 2026
Licence: Not confirmed: two readings disagree: DEEPSEEK LICENSE AGREEMENT and MIT License
Context window: 128,000 tokens — huggingface.co, read 16 October 2026
Longest answer: Not confirmed: The maker's pages reviewed do not print the standard maximum output-token limit for DeepSeek-V3.
Knowledge cutoff: Not confirmed: The maker's pages reviewed do not state a knowledge-cutoff date for DeepSeek-V3.
API price, input: Not confirmed: The current API pricing page does not state an input price for this exact model. The maker's DeepSeek-V3 announcement gives historical API pricing, not a current standard-tier price for DeepSeek-V3.
API price, output: Not confirmed: The current API pricing page does not state an output price for this exact model. The maker's DeepSeek-V3 announcement gives historical API pricing, not a current standard-tier price for DeepSeek-V3.
Where: Not confirmed: two readings disagree: weights and api, weights, app
Parameters: 671,000,000,000 parameters — huggingface.co, read 16 October 2026
Model card: https://huggingface.co/deepseek-ai/DeepSeek-V3 — huggingface.co, read 16 October 2026

Benchmarks

Measured by others, not by the maker: each score with the name the measurer gives the model and the setting it ran, its day and its licence.

Arena

Text Arena, deepseek-v3: 1,358 (1,354–1,363, 21,770 votes), published 15 October 2026

Arena’s leaderboard dataset (lmarena-ai/leaderboard-dataset on Hugging Face), CC BY 4.0: ratings from people’s votes between two anonymous models, as Arena publishes them. Data · CC BY 4.0

Epoch AI

Epoch Capabilities Index, DeepSeek-V3 (Mar 2025): 135.9 (132.5–137.8), as of 19 October 2026
Epoch Capabilities Index, DeepSeek-V3: 132.3 (127.5–135.9), as of 19 October 2026
GPQA Diamond, DeepSeek-V3-0324: 67.6% (± 2.7), run 1 April 2025
GPQA Diamond, DeepSeek-V3: 56.5% (± 2.8), run 27 January 2025
OTIS Mock AIME 2024–2025, DeepSeek-V3-0324: 37.8% (± 6.2), run 1 April 2025
OTIS Mock AIME 2024–2025, DeepSeek-V3: 15.8% (± 4.3), run 25 February 2025

Epoch AI, ‘Capabilities & benchmarking’, epoch.ai/benchmarks, CC BY 4.0: Epoch AI’s own runs and its Capabilities Index. Data · CC BY 4.0

Questions

Is DeepSeek-V3 still available?

Yes: it is generally available, as DeepSeek AI states it.

What is the context window of DeepSeek-V3?

The context window of DeepSeek-V3 is 128,000 tokens, as DeepSeek AI prints it.

Are the weights of DeepSeek-V3 open?

Yes: DeepSeek AI publishes the weights of DeepSeek-V3 for download.

What does DeepSeek-V3 take and return?

DeepSeek-V3 takes text and returns text.

How many parameters does DeepSeek-V3 have?

DeepSeek-V3 has 671,000,000,000 parameters parameters, as DeepSeek AI publishes it.

Sources

GitHub - deepseek-ai/DeepSeek-V3 (read 16 October 2026)
Introducing DeepSeek-V3 (read 16 October 2026)
Change Log (read 16 October 2026)
DeepSeek-V3/LICENSE-MODEL (read 16 October 2026)
deepseek-ai/DeepSeek-V3 (read 16 October 2026)

All of the Model Tracker