Sakana AI, an enterprise agent that generates analysis experiences of as much as 100 pages with slides, commercializes AB-MCTS with Sakana Marlin

Tokyo-based Sakana Ai shipped its first business product, the Sakana Marlin, this week. The Sakana group positions him as a digital CSO (chief technique officer). It is a B2B autonomous analysis agent constructed for enterprises.

Merlin would not reply in seconds like a chatbot. Give one analysis matter. It then operates autonomously for as much as about 8 hours. Every run returns an extended report and presentation slide deck. Based on Sakana, a whole lot to 1000’s of LLM queries are issued in a single session.

What’s fish marlin?

Merlin is a company analysis agent, not a chat assistant. Give one matter or query. Subsequent, plan your speculation, seek the advice of your sources, and independently confirm your outcomes. Weeks of technique work are compressed into hours.

Deliverables are structured for determination makers. Japan’s announcement features a report spanning dozens of pages. The English presentation cites a report that may be as much as about 100 pages lengthy. On the press convention, the report spanned 60 to 100 pages and cited 60 to 80 sources. Every report consists of the primary textual content, references, and appendices. Presentation slides are generated utilizing picture technology AI.

The Sakana group improved Marlin by a closed beta in April 2026. Through the beta, round 300 consultants examined Marlin on real-world duties. These duties ranged from technique improvement, market analysis, danger evaluation, and aggressive evaluation. Sakana additionally has a partnership with MUFG and has obtained strategic funding from Citigroup.

Inside AB-MCTS: wider or deeper?

The spine of Marlin is AB-MCTS (Adaptive Branching Monte Carlo Tree Search). It comes from previous analysis on fish.”Is it wider or deeper? Scaling LLM inference time calculations with adaptive branching tree search”

AB-MCTS treats inference as a tree search drawback. At every step, the algorithm makes one determination. You may additional develop your scope by producing new reply strategies. Or you may dig deeper by enhancing on a promising present reply. Customary iterative sampling simply scales up in parallel and hopes one reply is right.

The Multi LLM variant provides a second alternative. You may route steps to utterly completely different fashions. This collaboration was instrumental within the ARC-AGI-2 experiment reported by Sakana. The mixture of o4-mini, Gemini 2.5 Professional, and DeepSeek-R1 solved roughly 27.5% of the duties. The o4-mini mannequin alone solved about 23% of the issue. Marlin applies the identical adaptive search to longitudinal analysis.

The second essential part of Marlin is workflow automation. Fish AI Scientist Project. This mission demonstrated autonomous scientific discovery and was printed in Nature.

Interactive demo: Embeddable widget (marlin-abmcts-demo.html) Present “broader or deeper” selections reside. Press “Run” and watch the tree develop. Inexperienced nodes have increased scores and spotlight the very best paths. Toggle ‘Multi-LLM’ to see steps routed between completely different fashions.

Merlin comparability

The Marlins compete on depth, not pace. Conventional investigative instruments can present solutions inside minutes to tens of minutes. Marlin deliberately spends hours to enhance the standard of its output. Competitor run instances beneath are approximate and reported, not official numbers.

The trade-off is evident. There might be longer wait instances and you’ll be charged for every run. In return, you get deeper speculation testing and a completed product. You may cancel a run at any time, however credit will proceed to be consumed.

Pricing

Sakana presents Professional, Group, and Enterprise ranges, in addition to pay-as-you-go pricing. Pay-as-you-go pricing begins at 100 credit per run, at $98 per credit score. Professional prices 150,000 yen per thirty days and consists of 2,000 credit. The group charge is 400,000 yen per thirty days, which incorporates 6,000 credit. Enterprise pricing is customized and consists of devoted help.

Utilization and examples

Merlin is appropriate for high-stakes questions the place analysis is the bottleneck. A selected instance extracted from the goal job is proven beneath.

• market entry: “Assessing the Japanese stablecoin and tokenized funds market following regulatory adjustments.” Marlin maps drivers, dangers, and structured choices within the report.
• danger evaluation: “Mannequin answer state of affairs for the closure of the Strait of Hormuz” Examine the hypotheses, not only a abstract, earlier than drawing a conclusion.
• Aggressive evaluation: Introducing 3 rivals and rating the distinction in place. You may be returned with slides prepared for technique evaluate.

Every instance suits one immediate and one unattended run. People nonetheless evaluate the cited output earlier than making a call.

Strive the engine your self: TreeQuest

Marlin can’t be self-hosted. Nevertheless, its core algorithms will be run right now. Sakana has open sourced AB-MCTS as TreeQuest beneath the Apache 2.0 license. Set up it, outline a technology operate, and run a set search price range.

import random
import treequest as tq

# Every node holds a user-defined state; rating have to be normalized to [0, 1].
def generate(parent_state):
    if parent_state is None:               # None means develop from the basis
        new_state = "Preliminary draft"
    else:
        new_state = f"Refined: {parent_state}"
    rating = random.random()                # swap this for an LLM-based rating
    return new_state, rating

algo = tq.ABMCTSA()                         # Adaptive Branching MCTS (variant A)
search_tree = algo.init_tree()

for _ in vary(10):                         # technology price range of 10
    search_tree = algo.step(search_tree, {"generate": generate})

best_state, best_score = tq.top_k(search_tree, algo, okay=1)[0]
print("BEST:", best_state, spherical(best_score, 3))

Swap out random scores from LLM judges to recreate real-world patterns. TreeQuest additionally comes with multi-LLM search and checkpointing capabilities for long-running runs. Checkpoints are essential as a result of API errors can happen in the midst of lengthy classes.

Benefits and drawbacks

Strengths

• Peer reviewed foundations: NeurIPS’ AB-MCTS and Nature’s AI Scientist.
• Accomplished deliverables akin to references, appendices, slides, and so on.
• Adaptive computing focuses its efforts on probably the most promising branches.
• An open supply core (TreeQuest) permits AI researchers to discover strategies.

Weak spot

• Lengthy execution instances make iterating slower in comparison with minute-by-minute investigation instruments.
• Automated experiences might include hard-to-find errors that require human evaluate.
• Pricing and design are geared toward companies quite than particular person builders.
• Merlin itself is closed. Solely the underlying algorithm is open.

Necessary factors

• Fish marlins carry out autonomous analysis for as much as about eight hours per job.
• A single run creates a report with dozens of pages and slides.
• It’s constructed on AB-MCTS (NeurIPS 2025 Highlight) and AI Scientist workflow (Nature).
• Participation pricing is pay-as-you-go. 100 credit per run, JPY 98 per credit score.
• Meant for finance, company technique, consulting, and suppose tank groups.

supply of data

• Fish AI — Fish Marlin Launch: https://sakana.ai/marlin-release/
• Fish AI — Fish Marlin Product web page: https://sakana.ai/marlin/
• Fish AI — AB-MCTS Analysis and TreeQuest: https://sakana.ai/ab-mcts/
• SayanaAI/treequest (GitHub, Apache 2.0): https://github.com/SakanaAI/treequest