Deep Cogito Cogito v2.1
by Deep Cogito
MIT-licensed 671B hybrid-reasoning open model with short reasoning chains, plus custom post-training on enterprise data
Cogito v2.1 is Deep Cogito's 671-billion-parameter open-weight hybrid reasoning model. It is released under the MIT licence and served by Together AI, Fireworks AI, Ollama and others. It suits enterprise AI platform teams that want a frontier-class open model they can self-host, or a lab that post-trains specialized models on their own proprietary data.
Deep Cogito is a San Francisco post-training research lab founded in 2024 by Drishan Arora and Dhruv Malrana, who helped build Google's AI Search products, AI Mode and AI Overviews. It published Cogito v1 Preview in April 2025, presenting Iterated Distillation and Amplification (IDA) as a self-improvement training method. The Cogito v2 preview followed in July 2025 at 70B, 109B MoE, 405B and 671B MoE sizes. The flagship, Cogito v2.1, shipped on 19 November 2025. It is a 671B mixture-of-experts model forked from the open-licensed DeepSeek-V3-Base and post-trained in-house with process supervision on its reasoning chains. The model has a hybrid reasoning switch (enable_thinking) that toggles between fast answers and step-by-step thinking. It supports single, parallel and multiple tool calls in both modes, has a 128K-token context window and ships under the MIT licence. The full-precision BF16 checkpoint is about 1.3 TB and needs at least 8 NVIDIA B200s or 16 H200s; an FP8 variant fits on 8 H200s, and it serves on vLLM and SGLang. Hosted endpoints run on Together AI ($1.25 per million tokens), Fireworks AI (dedicated on-demand GPUs), OpenRouter, Ollama Cloud, Baseten and RunPod, and Ollama lists more than 220,000 pulls. Deep Cogito claims v2.1 uses the fewest average tokens of any reasoning model of similar capability. It says Cogito v2's 671B model produced reasoning chains about 60% shorter than DeepSeek R1 0528, and that all eight Cogito models from 3B to 671B cost under $3.5 million combined to train. The company itself cautions that benchmarks do not fully capture real-world performance. On 26 August 2026 Deep Cogito raised a $43 million Series A led by TQ Ventures, with Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons and Zscaler, taking total funding above $56 million. Zscaler is also a customer: Deep Cogito post-trains specialized models on a company's proprietary data and outcomes.
The head of an enterprise AI platform team that must run a frontier-class reasoning model on its own infrastructure under a permissive licence, or wants a partner to post-train a specialized model on proprietary data.
A commercially usable, MIT-licensed 671B reasoning model with vendor-reported shorter reasoning chains, which lowers token spend per task compared with longer-thinking open models.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Free, Usage-based, Contact for pricing
- Target Market
- CTOs, Heads of AI, ML Platform Engineers, Enterprise Developers
- Deployment
- Open-source, Self-hosted, API-based
- Founded
- 2024
- Headquarters
- San Francisco, United States
Key Features
- ✓Hybrid reasoning toggle
One flag (enable_thinking) switches the same model between fast direct answers and step-by-step reasoning. Teams can trade latency for depth per request without deploying two models.
- ✓Short reasoning chains
Process supervision during post-training aims for fewer thinking tokens. Deep Cogito reports the lowest average token use among reasoning models of similar capability, which cuts cost per task.
- ✓Native tool calling
Supports single, parallel, multiple and parallel-multiple function calls in both standard and reasoning modes, so it can drive agent workflows directly.
- ✓MIT-licensed open weights
Full BF16 weights and an FP8 variant are on Hugging Face under the MIT licence, allowing commercial use, modification and fully private self-hosting.
- ✓Broad hosting ecosystem
Available through Together AI, Fireworks AI, OpenRouter, Ollama Cloud, Baseten and RunPod, and runs locally with vLLM, SGLang, Ollama and Unsloth. This avoids single-provider lock-in.
- ✓Enterprise post-training
Deep Cogito builds specialized models post-trained on a company's own data and outcome metrics, going deeper than lightweight fine-tuning, as in its work with Zscaler.
Capabilities
Use Cases
- •Private reasoning assistant
A regulated enterprise serves the FP8 checkpoint on eight H200s inside its data centre, so sensitive analysis never leaves its own infrastructure.
- •Tool-using agent backend
A platform team uses parallel tool calling with reasoning switched on for complex steps and off for routine lookups, keeping latency and token spend under control.
- •Domain-specialized security model
A security vendor works with Deep Cogito to post-train a model on its product data and target metrics, as Zscaler did before investing.
- •Low-cost batch reasoning
A data team runs large volumes of document reasoning through Together AI's serverless endpoint at $1.25 per million tokens instead of paying closed frontier-model rates.
Ideal For
Best For
- ✓Self-hosting a frontier-class reasoning model inside a private data centre or VPC under an MIT licence
- ✓Agentic workflows that need native parallel tool calling with a switchable reasoning mode
- ✓Cost-sensitive reasoning workloads served through Together AI's serverless endpoint at $1.25 per million tokens
- ✓Enterprises that want a specialized model post-trained on proprietary data and business outcomes, as Zscaler did
- ✓Platform teams benchmarking open-weight alternatives to closed frontier models before committing to a vendor
Not Ideal For
- ✗Teams without multi-GPU Blackwell or Hopper capacity that still need on-premise deployment. Full precision needs 8 B200s or 16 H200s, so smaller open models are more practical.
- ✗Organisations whose policies exclude DeepSeek-derived weights. Cogito v2.1 is post-trained from DeepSeek-V3-Base despite its 'best US open model' positioning.
- ✗Buyers who need a first-party commercial API with vendor SLAs. Hosted access is through third-party inference providers, and enterprise post-training is sold by engagement with no published pricing.
- ✗Workloads that need million-token context. The model card specifies 128K tokens.
Deployment
Market Analysis
Pros
- ✓MIT licence permits unrestricted commercial use and fully private deployment
- ✓Vendor-reported token efficiency, with chains about 60% shorter than DeepSeek R1 0528 for the v2 671B model, which lowers inference cost per task
- ✓Hybrid reasoning and parallel tool calling in one checkpoint suit agent workloads
- ✓Wide hosting availability, with more than 220,000 pulls on Ollama
- ✓Backed by Benchmark and TQ Ventures, with a named enterprise customer (Zscaler) validating the post-training offer
Cons
- ✗Self-hosting is heavy: the BF16 checkpoint is about 1.3 TB and needs 8 B200s or 16 H200s, and FP8 still needs 8 H200s
- ✗No new open-weight release appears on Deep Cogito's research page since Cogito v2.1 in November 2025
- ✗Benchmarks are vendor-reported, and the company says they do not fully capture real-world performance
- ✗Built on DeepSeek-V3-Base, which may conflict with procurement policies that exclude DeepSeek-derived models
- ✗Hosted access depends on third parties; Fireworks lists no serverless option and no fine-tuning for this model
- ✗Thin independent practitioner coverage: Hacker News launch threads drew only a handful of comments
Pricing
Open weights (Hugging Face)
$0
- ✓MIT licence
- ✓BF16 and FP8 checkpoints
- ✓Self-host on vLLM, SGLang or Ollama
Together AI serverless API
$1.25 per 1M tokens (input and output)
- ✓OpenAI-compatible API
- ✓Hybrid reasoning
- ✓Native tool calling
Enterprise post-training
Contact for pricing
- ✓Specialized models post-trained on proprietary data
- ✓Optimisation against customer-defined metrics
The weights are free under MIT, so self-hosting costs are GPU costs: at least 8 B200s or 16 H200s at full precision, or 8 H200s with FP8. Together AI charges $1.25 per million input and output tokens. Fireworks AI offers only dedicated on-demand GPUs for this model (no serverless option) and does not show a per-token price. Deep Cogito publishes no pricing for its enterprise post-training work.
Security & Compliance
Connect
Sources
This page was written from 12 sources, 9 on domains other than deepcogito.com.
- 1.deepcogito.com — deepcogito.comvendor
- 2.deepcogito.com — researchvendor
- 3.deepcogito.com — cogito v2 1vendor
- 4.huggingface.co — cogito 671b v2.1
- 5.huggingface.co — cogito v2 1
- 6.together.ai — cogito v2 1 671b
- 7.fireworks.ai — cogito 671b v2 p1
- 8.ollama.com — cogito 2.1
- 9.unite.ai — deep cogito raises 43m series a to build the post training e
- 10.stocktitan.net — deep cogito raises 43m series a to advance the post training
- 11.southparkcommons.com — deep cogito
- 12.hn.algolia.com — search
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Abacus.AI Smaug
Open-weight Smaug Agentic, Flash and Mini models fine-tuned for long-running enterprise AI agents
GPT-6 Astra
OpenAI's frontier model for autonomous computer use, gated cyber capability and long-horizon coding
Inkling
Apache-2.0 975B-parameter open-weights model from Mira Murati's lab, built to be fine-tuned rather than rented
Fundamental NEXUS
A foundation model built for tables, not text — enterprise prediction without feature engineering