Fortytwo bets on collective reasoning with Engine

Fortytwo launches Engine, an API that combines multiple AI models to improve Qwen3.8-27B responses, using nearly 25% fewer tokens.

Improving the output of an artificial intelligence model without modifying its weights or increasing the cost per token. This is the core principle of Fortytwo Engine, a new inference API that leverages multiple models to guide collective reasoning. For its launch, the company is offering Qwen3.8-27B, with reported scores increasing by an average of 7% across six benchmarks and a reduction in token consumption of nearly 25% on software engineering tasks. Multiple models to guide a single reasoning process

Fortytwo Engine relies on a technology named Swarm Inference, developed to enable several artificial intelligence systems to collaborate during the generation of a response.

Rather than letting a single model follow a sequence of steps on its own, the system leverages evaluations from other models to guide the reasoning process as it unfolds.

Participants can review different proposals and help select the directions deemed most relevant. The goal is notably to reduce unnecessary developments, errors that propagate from one step to the next, and long correction sequences.

This approach builds on Fortytwo's work regarding consensus among models. In its research paper Swarm Inference with Peer-Ranked Consensus, the team describes a mechanism where responses are compared with one another, then weighted based on the reputation and evaluations of the participants.

Engine translates this research into an inference service accessible via an API. However, the public documentation does not fully detail the internal mechanisms used for each commercial request. Qwen3.8-27B as the first model

The first model offered is Qwen3.8-27B, an open-weight model that can already be used independently of Fortytwo.

Thus, the company does not present Engine as a new foundational model. The service constitutes an inference layer that modifies how responses are produced and evaluated.

Developers continue to use Qwen3.8-27B, but their requests pass through Fortytwo's collective reasoning system.

The company announces an average improvement of approximately 7% across the six selected evaluations, with an unchanged price per token compared to its baseline offering.

Other models are expected to join the platform, with no specific timeline announced for their availability. Reported gains across six evaluations

Fortytwo compares Qwen3.8-27B running standalone against its version running through Engine, under identical evaluation conditions.

On DeepSWE, dedicated to software engineering tasks, the score rises from 42.2 to 47.8, representing a 5.6-point gain.

The model also progresses across several reasoning and knowledge tests.

Évaluation · Base model · Fortytwo Engine DeepSWE · 42,2 · 47,8 GDP.pdf · 18,5 · 20,6 GPQA Diamond · 90,9 · 95,1 Humanity's Last Exam · 34,7 · 35,8 LiveCodeBench v6 · 80,2 · 84,1 SciCode · 46,8 · 49,0

These results show improvements ranging from 1.1 to 5.6 points depending on the tasks. The largest relative improvement is on DeepSWE, with an increase of roughly 13.3%.

The figures come from the evaluations published by Fortytwo. They do not constitute an independent validation of the system's superiority across all use cases.

The comparison also focuses on a specific model, Qwen3.8-27B. Nothing yet indicates that the same gains would be achieved with all open models. Nearly 25% fewer tokens on DeepSWE

One of the most interesting results concerns the volume of tokens required to complete a task.

Fortytwo compared the two configurations across 450 runs of DeepSWE.

The standalone model consumes an average of 16.3 million tokens per run, compared to 12.3 million with Engine, representing a 24.7% reduction.

This difference can impact the costs of agentic tasks, where multiple model calls and long reasoning sequences rapidly increase consumption.

The mechanism aims to prevent a model from continuing too long down a wrong path before back-tracking on its reasoning.

However, the reported gain applies specifically to this software engineering setup. It does not guarantee an identical reduction in a simple chat conversation, document analysis, or a task leveraging other tools.

Fortytwo also claims to maintain a speed comparable to that of the baseline model, without publishing detailed latency measurements in this presentation to verify this equivalence across different workloads. An API compatible with existing tools

Fortytwo Engine is accessible through an API compatible with OpenAI's Chat Completions format.

Developers have an endpoint to list models and another to send their conversations.

The service supports real-time streaming responses, function calling, and structured outputs.

For Qwen3.8-27B, it accepts text and images as input, with text output. Its context window reaches 262,144 tokens, while the stated output generation limit is 235,929 tokens.

This compatibility aims to facilitate its integration into assistants, development tools, and applications already using language model APIs.

It does not mean, however, that all parameters and behaviors from competing providers are replicated identically. Pricing starting at $0.425 per million tokens

Fortytwo lists the following rates for Qwen3.8-27B:

Usage · PPrice per million tokens Input · 0,425 $ Output · 2,55 $ Cached input · 0,085 $

The platform offers several access tiers with distinct rate limits.

The Core offer allows 20 requests per minute and 500 per hour. Pro increases to 200 requests per minute and 7,000 per hour, while Scale reaches 600 requests per minute and 20,000 per hour.

Fortytwo also announces a default zero data retention policy for requests processed via its API.

The company presents these terms as a way to protect the information entrusted to the service. Technical and contractual conditions remain to be examined according to the privacy requirements of each organization. Infrastructure stemming from research on model networks

Fortytwo has been developing an approach to inference based on model collaboration for several years.

Its work on weighted consensus notably