Talos-1: a production model for agentic work in European companies

Today we are releasing Talos-1. It started as a proof of concept for a new model architecture and ships as a production model for agentic work in European companies.

Over the past two and a half months we gave preview builds to beta testers in banking, telecommunications, education, accounting and other industries. Their feedback shaped the model we release today. It is the best model built in Europe.

Talos-1 is built for long, tool-heavy work inside the systems a company already runs. That focus shows in the benchmarks. On agentic tasks it scores level with models many times its size. On repository-scale coding it trails the frontier, which we did not optimise for.

Talos-1 ships with three reasoning efforts (low, medium and high) and a 200k context window.

What we built it for

The work European companies bring us is agentic: many steps, many tools, one outcome that either holds up or does not. So we spent the training budget there.

Beta testers ran Talos-1 in far more scenarios than we can list. Four of them: A bank used it as the engine behind internal agents that work across its core systems. A telecommunications group put it behind a search bot that answers staff questions from internal documentation. A learning platform let it read student homework and write individual feedback for each submission. An accounting team had it extract line items from receipts and book them under the rules of the country the receipt came from.

These four look different and share one shape. Each one needs the model to follow a long instruction without drifting, to call a tool and use what comes back and to read a document that was never written for a machine. Two things follow: the model has to be cheap to run on long inputs and it has to score on agentic benchmarks. We take them in that order.

Two cost curves

Talos-1 runs on an architecture we designed ourselves, from the attention layers up. We built it for one reason, the cost of long context. It delivered a second result we did not plan for: stronger reasoning. The cost side shows in two curves. We host the model ourselves, so those curves set what we can offer: how much context, at what price, to how many customers at once.

A language model does two kinds of work: it reads and it writes. Reading covers both an input at inference and a sequence during training. It is arithmetic. Every token is processed against the tokens before it. On a conventional transformer that cost grows with the square of the length. A training step is the same computation run forward and then backward, so reading and training share one curve. Writing is different. The model produces one token at a time. For each one it pulls the conversation so far out of GPU memory. On current hardware the limit for that step is memory bandwidth. Two kinds of work, two curves.

The comparison throughout is a conventional full-attention MoE transformer with the same dimensions and the same active parameters, using the key-value cache layout of the widely deployed open MoE models of 2026: eight key-value heads of 128 dimensions. The figures are derived from the architecture. They count compute in FLOPs and memory traffic in bytes per token. They do not measure wall-clock time on any particular kernel. Anyone with the model configuration can recompute them. The curves run to one million tokens to show where the design leads.

Reading and training. At 4k both models spend about 7.5 GFLOPs per token. At 256k the conventional model spends 92 and Talos-1 11, an 8.2× gap. At 512k it is 11.4×, at one million 14.5×: 350 GFLOPs against 24. Below 16k the two are within ten percent of each other. The same ratios apply to training. A conventional model pays 12.2× more compute per token to train at 256k than at 4k; Talos-1 pays 1.6×. That is what makes long-context training affordable for a team our size.

Writing. Per generated token at six concurrent requests, Talos-1 moves 32× less data through memory than the conventional model at 256k, 53× at 512k and 79× at one million: 13 GB read per token against more than a terabyte.

Talos-1 is built for long work: a contract with its annexes, a day of support tickets, a multi-step task held in view from start to finish. That needs context at a cost that allows daily use.

How it measures up

Bold marks the best score in each row among the four comparison models. "text": the model takes no images. "-": no published score. Talos-1 ran at reasoning effort high with each benchmark's own harness; BrowseComp ran in our own harness with web search and Python. The comparison models ran at their highest effort; their scores come from the official leaderboards and the vendors' own reports. The full table on the model page has all 14 benchmarks, seven models, the harness for every row and every source.

The table confirms the focus. On Agents' Last Exam, Talos-1 scores 49.0, six points behind Claude Opus 5 and within three of Kimi K3. On IFBench, which tests how closely a model follows written instructions, it reaches 82.2, nine points ahead of the best comparison model in the table. On AutomationBench, Zapier's test of business workflows across 47 simulated SaaS tools, it completes 42.5 percent of tasks end to end, six points behind GLM-5.3. On Toolathlon it scores 74.4, level with GPT-5.6 Sol and a point above GLM-5.3. On BrowseComp it finds 86.7 percent of the answers, four behind Claude Opus 5.

Document work is a strength. On OmniDocBench, which measures how well a model parses layouts, tables and formulas in real documents, Talos-1 reaches 90.8, within two points of GPT-5.6 Sol. The accounting use case above is exactly this capability: a receipt is a document with a layout and the layout carries meaning.

Coding is where the gap is. Talos-1 trails Claude Opus 5 by 22 points on DeepSWE. That gap is a decision. Every training hour we did not spend on repository-scale coding went into tool use, instruction following and document handling, because that is what our customers run.

Terminal-Bench, NL2Repo, MMMU-Pro, GPQA Diamond and Humanity's Last Exam are on the model page. The reasoning gain from the architecture shows there on GPQA Diamond, where Talos-1 scores 90.1, within four points of the field. Otherwise the pattern holds: close on knowledge and documents, behind on coding and on Humanity's Last Exam, which rewards sheer size.

Early feedback on Talos-1

Partners ran preview builds on their own workloads over the summer. In their words:

Limits, and what comes next

The context window at launch is 200,000 tokens. That covers a long contract, a full day of support tickets or a codebase of moderate size. It does not cover a whole regulatory filing with attachments. The limit is not the model. Talos-1 supports one million tokens and our evaluations ran it at up to 512k. The limit is inference capacity: a team our size serves what its hardware can hold. We will widen the window in phases, to 512k later this year and to one million after that.

There is no self-service API today, for the same reason. It opens once the capacity is there.

Running it

Of the three reasoning efforts, low answers fast and is meant for classification, extraction and short replies where the model has seen the pattern a thousand times. Medium is the default for agent loops, where each step is small and there are many of them. High spends more tokens thinking before it answers.

Launch pricing is €1 per million input tokens, €4 per million output tokens and €0.10 per million cached input tokens.

Companies that want to run Talos-1 contact us through the website. You will get a reply within 24 hours. From there we set up a pilot on your own use case.

Reach out →

Specifications, pricing and the full benchmark sources are on the Talos-1 model page.

← All news