Deep Cogito
10 models across 1 family · Latest: Cogito v2.1 671B (2025-11)
Building general superintelligence through advanced reasoning and iterative self-improvement.
Deep Cogito's portfolio covers 10 active models across 1 current family, spanning coding, rag, and agents. Open a model detail page to compare provider routes and sourced benchmarks.
Covers 6 workload areas across 10 active tracked models; last verified 2026-06-29.
Use it for
- Teams evaluating coding, rag, and agents across this lab's releases
- Comparing model families before committing to a flagship
- Migration and pricing follow-ups across 10 tracked models
Do not use it for
- Choosing a hosting provider without opening a model page for price ladders
Active models
10
Current models from this lab, excluding deprecated ones
Active families
1
Current model families from this lab
Open catalog
10 open
0 open source / 10 open weights
Lowest output price
$0.100 /1M
Cheapest tracked output across active models, per 1M tokens
Latest dated release
2025-11-19
Cogito v2.1 671B
Freshness
2026-06-29
Researched 18d ago
Information
Release cadence
Showing 5 recent dated releases (full timeline below). Latest: Cogito v2.1 671B (2025-11-19).
Where this lab wins
- Coding: 1 tracked model with SWE-bench / HumanEval-style scores.
- RAG: 6 tracked models with ruler / needle retrieval benchmarks.
- Agentic: 10 tracked models with BFCL, tau-bench, and SWE-bench tool-use coverage.
- Long-context: 6 tracked models with context-token or InfiniteBench-class signal.
Flagship quality / price signal
Flagship: Cogito v1 Preview Llama 3B (best sourced coding quality-per-dollar in this portfolio).
Quality-per-dollar unavailable for this flagship — benchmark coverage or output token pricing is still missing.
Deep Cogito is an American AI research organization founded in 2024. Building general superintelligence through advanced reasoning and iterative self-improvement. Deep Cogito ships 1 model family totaling 10 models, with the most recent release Cogito v2.1 671B in 2025-11. Notable families include Cogito. Use it as a stable reference for lab background, release coverage, and follow-up model pages as they are added. View official API endpoints, benchmark performance, and coding/agent fit for every Deep Cogito model.
About
Deep Cogito is an AI research lab founded in June 2024 by Drishan Arora and Dhruv Malhotra. The company develops the Cogito series of hybrid open-weight models trained with Iterated Distillation and Amplification (IDA), enabling both direct and extended-thinking reasoning modes. Backed by South Park Commons. Located in San Francisco.
Featured models
| Model | Released | Context | Input price ($/1M) | Output price ($/1M) | License | Openness |
|---|---|---|---|---|---|---|
| Cogito v2.1 671B | 2025-11-19 | 128k | - | - | Llama 3 Community | Open weights |
| Cogito v2 Preview Llama 70B | 2025-07-31 | - | - | - | Llama 3 Community | Open weights |
| Cogito v2 Preview Llama 109B MoE | 2025-07-31 | - | - | - | Llama 3 Community | Open weights |
Model families
Recent releases
- Cogito v2.1 671B- 2025-11-19
- Cogito v2 Preview Llama 70B- 2025-07-31
- Cogito v2 Preview Llama 109B MoE- 2025-07-31
- Cogito v2 Preview Llama 405B- 2025-07-31
- Cogito v2 Preview DeepSeek 671B MoE- 2025-07-31
FAQ
Who founded Deep Cogito and when?
Deep Cogito was founded in 2024 and is associated with San Francisco, California, United States.
What models has Deep Cogito released?
Deep Cogito ships 10 models across 1 family: Cogito.
Is Deep Cogito's technology open source?
All tracked Deep Cogito models are open-weight or open-source.
Where is Deep Cogito headquartered?
Deep Cogito is headquartered in San Francisco, California, United States.
What is Deep Cogito known for?
Building general superintelligence through advanced reasoning and iterative self-improvement. Its most prominent tracked family is Cogito.
How can I access Deep Cogito's models?
Deep Cogito's models are available via Fireworks AI.
Explore related pages
Last reviewed: 2026-06-29. Data sourced from public lab announcements and provider documentation.