Epoch AI Notable AI Models — Parameters, Training Compute, Dataset Size, Hardware, Cost And Accessibility, 1950–2026

DatasetCSV.GZOpenAI ModelsEpoch AIMachine LearningTraining ComputeLLMResearchFree

1,055 notable AI models from 1950 to 2026 with developer, publication date, domain and task, parameter count, training compute in FLOP, training dataset and its size, Epoch's confidence rating, country, organisation type, notability criteria, citations, epochs, training time, hardware type and quantity, estimated training cost in 2023 USD, power draw, numerical format, and model, code and weight accessibility, one row per model, newest first, with multi-valued fields semicolon-separated and compute figures written out in full digits, from Epoch AI's Notable AI Models database as of 2026-09-06.

Use Cases

Compute-scaling and training-cost trend analysis
Frontier-lab and country share of notable models
Open-weights versus closed-model tracking
Hardware demand modelling from FLOP and chip counts

Methodology

Single fetch of Epoch's notable_ai_models.csv; abstract, author, notes and sub-15%-filled estimate columns dropped; headers snake_cased; comma lists rejoined with semicolons.

Update Schedule

Static snapshot. Epoch AI adds models continuously as they are released; refresh weekly.

Attribution

Source: Epoch AI, 'Data on AI Models' (epoch.ai/data/ai-models), CC BY 4.0.

Schema

name	type
model	string
organization	string
publication_date	date
domain	string
task	string
parameters	integer
training_compute_flop	integer
training_dataset	string
training_dataset_size	integer
confidence	string
country	string
organization_categorization	string
link	string
reference	string
citations	integer
notability_criteria	string
epochs	number
training_time_hours	number
training_hardware	string
hardware_quantity	integer

Showing 20 of 27 rows

Sample Data

link	task	model	domain	country	reference	confidence	parameters	organization	publication_date	open_model_weights	model_accessibility	notability_criteria	organization_categorization
	Language modeling/generation;Question answering	GPT-6 Astra	Multimodal;Language;Vision	United States of America	GPT-6 Astra	Confident		OpenAI	2026-09-03	false	API access	SOTA improvement;Discretionary	Industry
https://www.openevidence.com/blog/model-family	Language modeling/generation;Question answering;Medical diagnosis	OpenEvidence Darwin	Language;Medicine		Introducing the OpenEvidence Model Family	Confident		OpenEvidence	2026-09-03	false	API access	SOTA improvement
https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash	Language modeling/generation;Question answering;Image Understanding;Video understanding	Gemini 3.8 Flash	Language;Multimodal;Image generation;Video;Audio	United States of America	Gemini 3.8 Flash	Confident		Google DeepMind	2026-09-02	false	API access	Discretionary	Industry
https://z.ai/blog/glm-5.3		GLM-5.3	Language	China	GLM-5.3: Frontier Coding with Emergent Cyber Capabilities	Confident	744000000000	Z.ai (Zhipu AI)	2026-08-14	false	API access	Discretionary	Industry
https://openai.com/index/gpt-5-5-with-trusted-access-for-cyber/	Software vulnerability discovery and exploitation;Language modeling/generation;Question answering	GPT-5.5 Cyber	Cybersecurity	United States of America	Scaling Trusted Access for Cyber with GPT‑5.5 and GPT‑5.5‑Cyber	Confident		OpenAI	2026-08-11	false	API access	Discretionary	Industry

Get this via API

# 1. Add dAgentBase once, in any MCP client. No install, no vendor keys.
#    Claude.ai / Claude Desktop: Settings -> Connectors -> Add custom connector
#    Cursor / Claude Code / others: mcp.json
{
  "mcpServers": {
    "dagentbase": {
      "url": "https://dagentbase.com/api/mcp",
      "headers": { "Authorization": "Bearer dm_live_YOUR_KEY" }
    }
  }
}

# 2. Then ask your agent, in plain language:
#    "Preview 'Epoch AI Notable AI Models — Parameters, Training Compute, Dataset Size, Hardware, Cost and Accessibility, 1950–2026' and, if it fits, claim it and download the files."
#    Tools it will use: search_listings -> preview_listing -> purchase_listing -> get_download_urls

# 1. Add dAgentBase once, in any MCP client. No install, no vendor keys.
#    Claude.ai / Claude Desktop: Settings -> Connectors -> Add custom connector
#    Cursor / Claude Code / others: mcp.json
{
  "mcpServers": {
    "dagentbase": {
      "url": "https://dagentbase.com/api/mcp",
      "headers": { "Authorization": "Bearer dm_live_YOUR_KEY" }
    }
  }
}

# 2. Then ask your agent, in plain language:
#    "Preview 'Epoch AI Notable AI Models — Parameters, Training Compute, Dataset Size, Hardware, Cost and Accessibility, 1950–2026' and, if it fits, claim it and download the files."
#    Tools it will use: search_listings -> preview_listing -> purchase_listing -> get_download_urls

Free

one time · open license

Details

Date Range1950-07-02 — 2026-09-03

Rows1,055

Size91.5 KB

Files1

FormatCSV.GZ

Available formats

CSV.GZ91.5 KB

Freeopen