Introducing Celeris-1 Magnus. A model built for agentic work.

On τ³-bench banking, Magnus delivers 41.2% at a 55-second median, against GPT-5.6-sol’s 38.1% at 79 seconds.

It’s a hybrid diffusion model derived from qwen3.8-27b, optimized for agentic workloads.

For more info, check out: http://celeris.ai/celeris-1-magnus…

Celeris-1 Magnus: the fastest model for agentic use cases

Source: https://celeris.ai/celeris-1-magnus CELERIS-1 MAGNUS

Magnus is built for agents that need to think, use tools, and get things done, without waiting around.

AGENT · LIVEREAL TASK, REAL SPEED

PERFORMANCE

Highest solve rate in the field.

All 97 tasks of τ³-bench banking, head-to-head against gpt-5.6-sol, gpt-5.6-luna and gemini-3.7-flash, reasoning effort as published.

01 · τ³-BENCH BANKING

010203040100s80s60s50stime per task (p50) → slowertasks solved % →Celeris-1 Magnus55 s · 41.2% · @ lowgpt-5.6-sol79 s · 38.1%gemini-3.7-flash98 s · 32.0%Celeris-1 Magnus @ none49 s · 27.8%gpt-5.6-luna57 s · 18.6% **Figure 1.**Solve rate vs median time per task, 97 tasks, official grader. Up and to the left is better. Celeris-1 without agent training: 5.3%.

SOLVE RATE41.2%

VS GPT-5.6-SOL+3.1 pts

TIME PER TASK · P5055 s

NO AGENT TRAINING5.3%

Methodology →

Leads the hardest agent benchmark

41.2% on τ³-bench banking, ahead of gpt-5.6-sol, with the field’s best solve rate.

OFFLOW

A reasoning dial, per request

One flag between fast and thorough. 13.4 extra points when thinking is on, for 6 seconds.

Fast where it counts

55 seconds median per completed task, quicker than every model within ten points of it.

THE FAMILY

Celeris-1 or Magnus?

Celeris-1 is built for answers; Magnus is built for tasks. The API is the same, so switching is a one-word change.

Celeris-1

Ultra-fast generation for chat, search and real-time interfaces.

ModalityText

Context window131,072 tokens

Max output2,048 default

Best atInstant answers

Learn more

Celeris-1 Magnus

Agentic work: tool loops, long tasks, structured actions.

ModalityText + tools

Context window131,072 tokens

Max output16,384 tokens

Best atAgents that finish

Get an API key

GET STARTED

Start building.

The API is OpenAI-compatible. Point your SDK at inference.celeris.ai and keep the agent code you already wrote.

01

Create an account

Sign up at console.celeris.ai. No waitlist, no sales call.

02

Get an API key

Keys are live immediately and billed per token.

03

Point your SDK

Set the base URL and model id. Your existing agent code runs as is.

agent.pyPYTHON

from openai import OpenAI

client = OpenAI(
    base_url="https://inference.celeris.ai/celeris-1-magnus/v1",
    api_key="ck_…",
)

resp = client.chat.completions.create(
    model="celeris-1-magnus",
    messages=[{"role": "user", "content": "Move my Friday payment to Monday"}],
    tools=my_tools,
    # the reasoning dial: on when the task earns it
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)