Acree opens up new enterprise-focused, customizable AI model AFM-4.5B trained on ‘clean, rigorously filtered data’

Arcee.ai, a startup that focuses on developing small models of artificial intelligence for commercial use and institutions, Open its AFM-4.5B model for free use by small companies-Publishing weights on the face of embrace Allowing institutions with less than $ 1.75 million in annual revenue to use without fees under the “ACRE Model License”.

The parameter model of 4.5 billion of the parameters-which was designed to use institutions in the real world-is designed much smaller than tens of billions to trillion pioneering border models-between cost efficiency, organizational compliance and strong performance in a compressed fingerprint.

AFM-4.5B was one of two two parts by ACRE last month, and it is already “caught instructions”, or “Instructions” form, designed for chatting, retrieving and creative writing and can be immediately published for these cases of use in institutions. Another basic model has also been released at a time when it has not been seized on the instructions, just pre -trained, allowing more customization by customers. However, both were only available through the terms of commercial licensing – so far.

ACREE (CTO) Also note in a post on X more “Models dedicated to thinking and using tools on the road,” as well.

He wrote in another post: “The AFM-4.5B construction was a great effort, and we are very grateful to everyone who supported us that we cannot wait to know what it adopts.” “We just started. If you have notes or ideas, please do not hesitate to communicate at any time.”

The model is now available for publication through a variety of environments – from the cloud to smartphones to devices.

It is also directed towards the growing Acre list of institutions, needs and desires – specifically, a model that has been trained without violating intellectual property.

Acre also wrote in the first AFM-4.5b advertisement last month: “A huge effort was made to exclude books protected by copyright and materials with an unclear license.”

Acre notes that it has worked with Datologyai for an external data science party to apply technologies such as the source mixing, the comprehensive-based filter, quality control-and it is aimed at reducing hallucinations and IP risks.

Focus on the needs of the institution’s customers

AFM -4.5B is an arcee.ai response when you see the main pain points in the adoption of institutions from the Improvised Intelligence: High cost, limited allocation, and organizational concerns about the large royal language models (LLMS).

Over the past year, the Arcee team conducted discussions with more than 150 organizations, from startups to Fortune 100 companies, to understand the current LLMS restrictions and set their typical goals.

According to the company, many prevailing LLMS companies-such as those in Openai, anthropic, or Deepseek-expensive and difficult to customize the industry needs. Meanwhile, while models are smaller than open weight like Llama, Mistral and QWEN have provided more flexibility, fears about licensing, IP source and geopolitical risks.

AFM-4.5B has been developed as an “no trade” alternative: customizable, compatible, and costly effective without sacrificing the quality of the form or the ability to use.

AFM-4.5B is designed with the elasticity of publishing in mind. It can work in cloud, local, hybrid or even edge environments-thanks to their efficiency and compatibility with open work frameworks such as Ungging Face Transformers, Llama.CPP and (suspended version) VLLM.

The form supports quantum formats, allowing it to operate on lower graphics processing units or even central processing units, making it practical for applications with restricted resources.

The company’s vision secures support

The broader Arcee.ai strategy focuses on building small, adaptive language models Many cases are used within the same organization.

“You don’t need to go largely in business use,” said CEO Mark McCadad in an interview in the Venturebeat project last year. The company emphasizes rapid repetition and typical customization as the essence of its display.

This vision received the investor’s support with the A series A series A 24 million dollar in 2024.

Inside the architecture and training process in AFM-4.5B

The AFM-4.5B is used by the Decoder transformer structure only with many improvements to performance and elasticity of publishing.

It includes an assembled interest to inquire to infer faster and activate the Relu² instead of Swiglu to support contrast without insulting accuracy.

Training follows a three -stage approach:

• Premium pre -symbols 6.5 trillion of general data
• Training on 1.5 trillion symbol focuses on mathematics and symbol
• Adjusting the instructions using high -quality data groups, follow -up of instructions and learning to reinforce with verification and preference comments

To meet strict compliance criteria and intellectual property standards, the model has been trained on approximately 7 trillion data from data organized for safety of hygiene and licensing safety.

A competitive model, but not a leader

Despite its smaller size, AFM-4.5B performs competitively through a wide range of standards. The average version set on the instructions is 50.13 via the evaluation wings such as MMLU, Mixeval, Triviaqa, Agival-which excels over similar models such as GEMMA-3 4B-IT, QWEN3-4B and Smallm3-3B.

The multi -language test shows that the model offers a strong performance across more than 10 languages, including Arabic, Mandarin, German and Portuguese.

According to Arcee, adding additional dialects support is clear and direct due to its normative structure.

AFM-4.5B also showed a strong early jar in general evaluation environments. In the top -ranked forefront plate through the user’s voices and the rate of victory, the model is in general, as it only falls behind Claude Obus 4 and Gemini 2.5 Pro.

It is characterized by a 59.2 % victory rate and the fastest time in any higher model at 0.2 seconds, associated with a generation speed of 179 symbols per second.

Clear support for agents

In addition to general capabilities, AFM-4.5B comes with compact support to connect to jobs and thinking about the agent.

these The features aim to simplify the process of building artificial intelligence agents and tools to automate workflowReducing the need for complex engineering or coincidence layers.

This function corresponds to the broader Arcee strategy to enable companies to build custom production models faster, with a decrease in the cost of ownership (TCO) and the easiest integration in commercial operations.

What is the following for Acre?

AFM-4.5B represents Arcee.ai batch to determine a new category of ready -made language models for institutions: small, performance, and fully customized, Without settlements that often come with royal llms or open SLMS.

With competitive standards, multi -language support, strong compliance standards, and flexible publishing options, the model aims to meet the institution’s needs for speed, sovereignty and scale.

Whether Arcee can publish a always in the scene of changing artificial intelligence quickly dependent on its ability to fulfill this promise. But with AFM-4.5B, the company took a first confident step.