AI

Language models trained from scratch, more than two dozen others fine-tuned, and tools for working with AI agents.

Argonne base models
8
models on Hugging Face
48
datasets
47
model downloads
17,819
dataset downloads
40,139

Argonne: language models trained from scratch

Each point is a base model, placed by the date its Hugging Face repository was created and by its size; the larger the point, the more its family of checkpoints has been downloaded.

250M 500M 1B 2B 5B Jul 2025 Jan 2026 Jul 2026 Argonne 1.0: 276M parameters, on Hugging Face since March 3, 2025, 285 downloads across 2 checkpoints 1.0 276M Argonne 1.5: 357M parameters, on Hugging Face since March 14, 2025, 179 downloads across 1 checkpoints 1.5 357M Argonne 2.0: 4.92B parameters, on Hugging Face since January 24, 2026, 139 downloads across 1 checkpoints 2.0 4.92B Argonne 2.5: 1.27B parameters, on Hugging Face since April 5, 2026, 2,964 downloads across 5 checkpoints 2.5 1.27B Argonne 3.0: 2.88B parameters, on Hugging Face since June 1, 2026, 2,785 downloads across 3 checkpoints 3.0 2.88B Argonne 3.5: 2.88B parameters, on Hugging Face since August 1, 2026, 509 downloads across 2 checkpoints 3.5 2.88B Argonne 4.0: 1.04B parameters, on Hugging Face since August 13, 2026, 764 downloads across 2 checkpoints 4.0 1.04B Argonne 4.5: 2.06B parameters, on Hugging Face since September 29, 2026, 1,939 downloads across 5 checkpoints 4.5 2.06B

Training code on GitHub Every model on Hugging Face

Fine-tuned models

27 models built on other families' checkpoints, by what the fine-tuning taught them; each bar follows a model's downloads to date, on one scale.

  • Reasoning 10
  • Task models, 2024 8
  • Preference optimization 5
  • Tool calling 2
  • Format conversion 1
  • Chat 1

Reasoning

10 models

small models taught to think before they answer

  1. Argonne-Qwen1.5-0.5B-think from Qwen1.5-0.5B, Jul 2026 728
  2. Qwen3.5-0.8B-Opus-4.6-thinking from Qwen3.5-0.8B-thinking, Mar 2026 82
  3. Qwen3.5-0.8B-thinking from Qwen3.5-0.8B-Base, Mar 2026 182
  4. llama3.2-3b-thinking from Llama-3.2-3B, Mar 2026 138
  5. Qwen3-0.6b-thinking from Qwen3-0.6B, Feb 2026 1,524
  6. llama-3-2-1b-open-r1-mot-sft from Llama-3.2-1B, Nov 2025 84
  7. llama3.2-1b-thinking from Llama-3.2-1B, Nov 2025 2,339
  8. qwen2.5-0.5b-open-r1-mot-cot-sft from Qwen2.5-0.5B, Nov 2025 71
  9. finetuned-llama-3.2-3b-math-reasoning from Llama-3.2-3B, Nov 2025 40
  10. Qwen2.5-1.5B-Instruct-Lora-Deepseek-R1 from Qwen2.5-1.5B-Instruct, Mar 2025 100

Task models, 2024

8 models

question answering, sentiment, summaries and named entities

  1. roberta-large-ner from roberta-large, Jul 2024 393
  2. bert-base-ner from bert-base-cased, Jul 2024 57
  3. distilbert-base-cased-ner from distilbert-base-cased, Jul 2024 54
  4. t5-base-summary-model from t5-base, Jul 2024 46
  5. t5-large-summary-model from t5-large, Jul 2024 53
  6. finetuned_qa_squad_bert_base_model from bert-base-uncased, Jul 2024 275
  7. finetuned_sentiment_bert_base_model from bert-base-uncased, Jul 2024 1,174
  8. bert-base-qa-finetuned-squad-v2 from bert-base-uncased, Jun 2024 45

Curated and synthetic data

47 datasets on Hugging Face: cleaned pretraining text, answers and reasoning traces written by large models for training small ones, summaries, captions and preference pairs. Each bar follows a dataset's downloads to date, on one scale.

Which model wrote which

Most of the instruction sets come from one design: the same public prompts, answered by each of five large models. Darker cells were downloaded more.

OpenAssistantSynthetic promptsOpenR1 MathAlpacaDolly 15kUltraFeedback
Llama 4 Maverick 214 228 235 217 216
Llama 4 Scout 198 189 217 222 192
Llama 3.3 70B 162 175 202 209 237
Llama 3.3 8B 152 175 207 199 201
Mistral Large 217 148 235

Reasoning traces

13 datasets

worked solutions and chains of thought, for teaching models to think

  1. MiniMax-M2.1-Mixture-of-Thoughts 100K–1M rows, Dec 2025 5,753
  2. openmath-reasoning-medley 1M–10M rows, Feb 2026 4,648
  3. 0.5M-thinking 100K–1M rows, Jan 2026 4,243
  4. 0.9M-thinking 100K–1M rows, Jan 2026 4,057
  5. toucan-agentic-thinking 100K–1M rows, Jan 2026 1,016
  6. gsm8k-thinking 1K–10K rows, Dec 2025 947
  7. arxiv-qa-thinking 100K–1M rows, Jan 2026 863
  8. gsm8k-Llama-4-Maverick-17B-128E-Instruct-FP8 1K–10K rows, Jun 2025 464
All 13
  1. Llama-4-Maverick-17B-128E-Instruct-FP8-OpenR1-Math-220k 10K–100K rows, May 2025 235
  2. s1k-magistral-small-2506 1K–10K rows, Jun 2025 235
  3. Llama-4-Scout-17B-16E-Instruct-FP8-OpenR1-Math-220k 10K–100K rows, May 2025 217
  4. Llama-3.3-8B-Instruct-OpenR1-Math-220k 10K–100K rows, May 2025 207
  5. Llama-3.3-70B-Instruct-OpenR1-Math-220k 10K–100K rows, May 2025 202

Instruction responses

21 datasets

answers written by large models to public prompt sets, for fine-tuning small ones

  1. ProcessedOpenAssistant 10K–100K rows, Feb 2025 417
  2. mistral-awesome-chatgpt-prompts under 1K rows, Jun 2025 272
  3. Llama-3.3-70B-Instruct-databricks-dolly-15k 10K–100K rows, May 2025 237
  4. Llama-4-Maverick-17B-128E-Instruct-FP8-instruct-synthetic-prompt-responses 10K–100K rows, May 2025 228
  5. Llama-4-Scout-17B-16E-Instruct-FP8-alpaca-cleaned 10K–100K rows, May 2025 222
  6. ProcessedOpenAssistant-mistral-large-2411 10K–100K rows, Apr 2025 217
  7. Llama-4-Maverick-17B-128E-Instruct-FP8-alpaca-cleaned 10K–100K rows, May 2025 217
  8. Llama-4-Maverick-17B-128E-Instruct-FP8-databricks-dolly-15k 10K–100K rows, May 2025 216
All 21
  1. Llama-4-Maverick-17B-128E-Instruct-FP8-ProcessedOpenAssistant 10K–100K rows, May 2025 214
  2. Llama-3.3-70B-Instruct-alpaca-cleaned 10K–100K rows, May 2025 209
  3. Llama-3.3-8B-Instruct-databricks-dolly-15k 10K–100K rows, May 2025 201
  4. Llama-3.3-8B-Instruct-alpaca-cleaned 10K–100K rows, May 2025 199
  5. Llama-4-Scout-17B-16E-Instruct-FP8-Instruct-FP8-ProcessedOpenAssistant 10K–100K rows, May 2025 198
  6. Llama-4-Scout-17B-16E-Instruct-FP8-databricks-dolly-15k 10K–100K rows, May 2025 192
  7. Llama-4-Scout-17B-16E-Instruct-FP8-instruct-synthetic-prompt-responses 10K–100K rows, May 2025 189
  8. Llama-3.3-70B-Instruct-instruct-synthetic-prompt-responses 10K–100K rows, May 2025 175
  9. Llama-3.3-8B-Instruct-instruct-synthetic-prompt-responses 10K–100K rows, May 2025 175
  10. Llama-3.3-70B-Instruct-ProcessedOpenAssistant 10K–100K rows, May 2025 162
  11. Llama-3.3-8B-Instruct-ProcessedOpenAssistant 10K–100K rows, May 2025 152
  12. llama-awesome-chatgpt-prompts under 1K rows, Jun 2025 152
  13. instruct-synthetic-prompt-responses-mistral-large-2411 10K–100K rows, Apr 2025 148

Pretraining text

1 dataset

web text cleaned for training from scratch

  1. processed-fineweb-edu 1M–10M rows, Feb 2025 1,215

Code

1 dataset

seeds for generating programming data

  1. oss-code-seeds 100K–1M rows, Feb 2026 220

Data that trained the models here

As the models' own cards declare them.

MiniMax-M2.1-Mixture-of-Thoughts
0.5M-thinking

Every model, on one calendar

One circle per model, by the date its repository was created; the area follows its downloads to date.

Qwen fine-tunes 9
Argonne, trained from scratch 21
Llama fine-tunes 7
Nemotron 30B and tool-calling adapters 3
BERT, RoBERTa and T5 fine-tunes 8

Agents and tools

Sage

Ask your docs a question. Get an answer that links the exact section it came from.

Started
December 2025

GitHub

claudekeeper

Schedule prompts for Claude Code, and keep your sessions running through the 5-hour usage cap.

Started
July 2026
Version
0.3.1, 4 releases

GitHub PyPI

RedLLM

A GPT-style model pretrained on the classic novel Dream of the Red Chamber (红楼梦).

Started
February 2025

GitHub

hpcpilot, an agent for Slurm clusters, is on the HPC page.

Articles on language models and agents

  1. Evaluating AI Agents When Every Run Differs: Repeated Trials, Paired Comparisons and Graders
  2. Tool Calling in Language Model Agents: Designing the Interface Between a Model and Its Tools
  3. Workflow or Agent: Deciding Which Steps Belong in Code and Which Belong to the Model
  4. Why Temperature Zero Is Not Reproducible: Floating-Point Arithmetic and Batching in Language Model Inference
  5. Tuning the GPU Interconnect in Multi-Node Language Model Pretraining
  6. A Desktop Grace Blackwell Machine as Research Infrastructure: Serving and Training Benchmarks for the NVIDIA DGX Spark
  7. Evaluating an Agentic System: Three Axes, and What Auditing the Measurements Changed
  8. Argonne 3.5-base: Retraining an Unchanged Architecture With a Revised Recipe
  9. Argonne 3.5-think: From Base Model to Reasoning Model
  10. Pretraining a Language Model From Scratch: Argonne 1.0 to 3.0
  11. How I Taught a Small Language Model to Reason

Figures as of October 28, 2026: models, sizes and download counts from Hugging Face; Argonne sizes for 1.0 to 3.0 as the pretraining article reports them.