I.
Specialists beat generalists.
The next decade of AI is about depth, not breadth. A model post-trained on a real distribution will out-reason a generalist chat model an order of magnitude larger — at a tenth the cost. We build the specialists.
Frontier AI research lab
Agents that reason, collaborate, and deliver.
Macrodeep is a frontier AI research lab. We train our own specialized industry models and build the autonomous agents, platform, and tools that run on them — so every team can ship pipeline, code, and decisions at a scale no team has touched before.
Autonomous sales intelligence◆Fleet-scale coding agents◆Frontier GTM foundation model◆Trust-gated execution◆Human-in-the-loop by design◆
Autonomous sales intelligence◆Fleet-scale coding agents◆Frontier GTM foundation model◆Trust-gated execution◆Human-in-the-loop by design◆
0+
specialized AI agents, shared brain
1M
context window on Nico 2.5
0M+
real GTM training trajectories
0x
average pipeline lift in month one
What we believe
Three things most labs still get wrong.
II.
Reasoning is not the same as completion.
Most models guess. We train ours to plan, act, and revise — against verifiable outcomes, not preference ratings. Replies sent. Meetings booked. Deals won. Code that compiles and ships.
III.
Autonomy belongs in the work, not the workflow.
The future is not chat. It is autonomous systems that own outcomes — research, draft, send, follow up, update the CRM, log the call, escalate to a human only when trust gates demand it.
Building the first artificial general agents.
Autonomous AI will reshape every industry. Macrodeep is building the agents — and the infrastructure to run them — so every business can operate at a scale that was never possible before.
Macrodeep Research
Frontier research, shipped as product.
We train our own specialized industry models and the agentic systems that run on them. Our research moves straight from the lab into production — no middlemen, no generalist chat wrappers.
Featured · March 8, 2026
Introducing Nico 2.5 — a frontier model, built for revenue.
A 120B-parameter mixture-of-experts post-trained on real outbound, real replies, real deals — paired with live retrieval over our knowledge graph at inference. The result is a model that doesn't guess about your buyers; it reasons about them, with citations.
Read the paperFoundation models
Specialist beats generalist.
We post-train frontier-class models on the exact distribution they're asked to serve — real revenue artifacts, not the open web. Sharper, smaller, cheaper.
Agentic reasoning
Models that plan, act, and revise.
Multi-stage post-training against verifiable task outcomes — replies, meetings booked, deals won. Our models don't just complete; they execute multi-step workflows.
Multi-agent systems
A fleet with one brain.
Twenty specialized agents sharing a typed memory graph. Cross-channel signal aggregation and trust propagation produce coherent behavior across months of deal history.
Synthetic data
Data at the edge of the frontier.
When supervised data runs out, we generate it — sandboxed agentic trajectories at a scale no crawl can match. The next frontier is data, not compute.
Model releases
In production today.
Nico 2.5GTM
1M context · retrieval-native ProductionNico CodeCoding
agentic SWE-bench 76–82% RoadmapOlivia v3Conversation
grounded dialog · citations ProductionCortexReasoning
trust-gated action policy ProductionField notes
From the lab, this week.
Nico 2.5 reaches 87% reply-rate uplift over Claude Sonnet on the internal outbound eval (n=12,000).
Salestools GTM agent fleet processed 480K autonomous CRM writes across design partners last week.
Cortex routing live in production for 12 enterprise customers — trust-gated execution, full audit trail.
Grain ships native macOS desktop app with bundled local executor — zero SaaS lock-in.
Nico 2.5 architecture, training stages, and corpus published. Read the paper.
Training-data-at-scale post: how we curated the corpora behind Nico 2.5.
Evaluations
A specialist beats a generalist on every revenue task we measured.
Internal head-to-head benchmarks. Same prompts, same accounts, same scoring rubrics. Numbers refresh per release.
Task
Nico 2.5
GPT-4o
Claude Sonnet
First-touch reply rate (B2B outbound, n=12K)
8.4%
4.1%
4.6%
Account brief — citation accuracy
94%
71%
78%
Stakeholder identification @top-3
0.81
0.62
0.65
Multi-step pipeline workflow (success)
76%
48%
52%
Cost per million output tokens (USD)
$0.30
$10.00
$15.00
Methodology and full results in the Nico 2.5 model card.
Products
A full stack for autonomous work.
One mission, four surfaces. The flagship agent, the platform it runs in, the model that powers them, and the autonomous sales intelligence that feeds them.
Flagship · GTM Agent
Salestools GTM.
An autonomous revenue teammate, not another chat wrapper.
It researches each account, personalizes every touch, enriches pipeline, and runs multi-step GTM workflows inside the CRM and inbox you already use — with humans in the loop where trust requires it.
Explore the agentSalestools
Twenty specialized agents research accounts, score leads, run cross-channel outreach, and heal pipeline end-to-end.
Learn more
Grain
The command center for teams running AI agents. Review, approve, and ship work from one surface — web, desktop, or mobile.
Learn moreNico 2.5
Frontier go-to-market foundation model. 120B active, 1M context, retrieval-native — the brain behind everything.
Learn morePrinciples
How we build.
Three commitments that shape every product decision.
01
Autonomy with oversight
Trust-gated execution. Agents do the work; humans approve what matters. Every action is logged, replayable, and reversible.
02
Grounded reasoning
No hallucinated claims. Every output ships with a citation trail — CRM fields, research sources, call transcripts, prior deal history.
03
Privacy by default
Your CRM and prospect data is never used to train models. Aggregate, opt-in telemetry only — and enterprise deployments can run entirely inside your network.
Start with the agents.
Deploy an autonomous team for sales or engineering in the time it takes to make a coffee.