Arcee.AiStartup focusing on developing small AI models for industrial and enterprise, IS Openly Own AFM-4.5B for limited free use by small companies-public Weights on hugging the face and permission to enter enterprises that earn lower than $ 1.75 million of annual revenues for using it without fees under Non -standard “ACREE model license.“
Parameter model value 4.5 billion parameters in the real world-significantly smaller than tens of billions to billions of leading border models-cost-effectiveness, regulatory compatibility and good performance in a compact trace.
AFM-4.5B was One of the two parts of the Acree release last monthAnd there is already a “tuned manual” or the “instruction” model, which is intended for chat, downloading and creative writing and might be implemented immediately for these cases of use in enterprises. Another basic model was also released at a time that has not been tuned, but pre -trained, enabling greater configurability by customers. However, each were only available on the basis of industrial licensing conditions – so far.
ACREE (CTO) technology director Lucas Atkins also recorded in Post on X that is more “Dedicated models of reasoning and using tools are also on the road.”
The AI Impact series returns to San Francisco – August 5
The next AI phase is here – are you ready? Join the leaders from Block, GSK and SAP to see the exclusive look at how autonomous agents transform the flows of the work of the company-decision-making in real time for comprehensive automation.
Secure your house now – the space is limited: https://bit.ly/3guplf
“Building AFM-4.5B was a huge team effort and we are deeply grateful to everyone who supported us, we can’t wait to see what you are building with him,” he He wrote in one other post. “We’re just starting. If you have opinions or ideas, don’t hesitate to reach out at any time.”
The model is now available for implementation in various environments – from clouds to smartphones to Edge equipment.
It is also focused on the growing list of business clients and their needs and needs – in particular the model trained without violation of mental property.
How Acree wrote in his first post AFM-4.5B last month: “He put a huge effort to exclude books and materials protected by copyright with unclear licensing.”
Acree notes that he has collaborated with the curative company of other corporations Datologyai To apply techniques akin to mixing sources, filtering based on embedding and quality control-all aimed at minimizing hallucinations and IP risk.
Focused on the needs of the company’s clients
AFM-4.5B is the ARCEE.AI response to what it considers to be the essential pain points in the party’s party AI: high cost, limited configuration and regulatory fears regarding reserved models of large languages (LLM).
Over the past 12 months, the ARCEE team has discussed with over 150 organizations, from startups to the company from the Fortune 100 list to know the limitations of existing LLM and define their very own model goals.
According to the company, many corporations have found the essential LLM-as of those from OpenAI, Antropic or Deepseek-too expensive and difficult to adapt to the needs in the industry. Meanwhile, while smaller open models, akin to Llama, Mistral and Qwen, offered greater flexibility, introduced concerns about licensing, IP origin and geopolitical risk.
AFM-4.5B was developed as an alternative “without trade”: configurable, consistent and profitable without sacrificing the quality and usability of the model.
AFM-4.5B has been designed with the flexibility of implementation. It can work in cloud, local, hybrid and even edges-thanks to its efficiency and compatibility with open frames, akin to hugging face transformers, llama.cpp and (waiting) vllm.
The model supports quantized formats, enabling the GPU to work with a lower frame and even processors, which makes it practical for applications with limited resources.
The company’s vision protects support
The wider strategy of Arcee.Ai focuses on building small languages models that may supply the Small Language models (SLM) Many cases of use in the same organization.
As explained by the general director of Mark Mcquade in an interview with Venturebeat last 12 months: “You don’t have to go so large for business use.” The company puts emphasis on fast iteration and adaptation of the model as the basis of its offer.
This vision gained the investor’s support due to the $ 24 million round in 2024.
Inside the architecture and training AFM-4.5B
The AFM-4.5B model uses the architecture of the transformer only a decoder with several optimizations regarding the performance and flexibility of implementation.
It accommodates the grouped attention of the query for faster inference and activation of Relue² as a substitute of SWIGLI in order to handle spraceification without reducing accuracy.
The training took place in a three -phase approach:
- Conductivity on 6.5 trillions of general data tokens
- Connection on 1.5 trillion tokens emphasizing mathematics and code
- Tuning instructions using prime quality data sets for instructions and learning to strengthen with verifiable and preference feedback
To meet strict conformity and IP standards, the model has been trained on almost 7 trillions of data tokens chosen for purity and licensing security.
Competitive model, but not a leader
Despite the smaller Size AFM-4.5B, it really works competitively in a wide range of comparative tests. The tuned version to the instructions is a result of 50.13 in evaluation apartments akin to MMLU, Mixeval, Triviaqa and Algival-connecting or exceeding models of similar size, akin to GEMMA-3 4B-IT, QWEN3-4B and SMOLM3-3B.
Multilingual tests show that the model provides good results in over 10 languages, including Arab, Mandarin, German and Portuguese.
According to Arcee, adding support for additional dialects is easy on account of its modular architecture.
AFM-4.5B also showed a strong early adhesion in public assessment environments. In the leader board, which assesses the quality of the conversation model in line with the voices of users and the winning indicator, the model ranks third in the general classification, raising only Claude Opus 4 and Gemini 2.5 Pro.
It offers a win indicator of 59.2% and the fastest delay in any best model after 0.2 seconds, combined with a generation speed of 179 tokens per second.
Built -in support for agents
In addition to the general capabilities of AFM-4.5B, it is equipped with built-in service for calling functions and agency reasoning.
These The functions are aimed at simplifying the process of building AI agents and tools for automating work flowReducing the need for complex fast layers of engineering or orchestration.
This functionality is in line with the broader Arcee strategy consisting in enabling enterprises to build non -standard models ready for production, with lower total ownership (TCO) and easier integration with business operations.
What next with Acree?
AFM-4.5B represents ARCEE.Ai to strive to define a latest category of language models ready for a company: small, efficient and configurable, Without compromises, which are often associated with reserved LLM or SLM open weight.
Thanks to competitive comparative tests, multilingual support, strong conformity standards and flexible implementation options, the model goals to fulfill the company’s needs in terms of speed, sovereignty and scale.
Whether Arcee can detect a everlasting role in the rapidly changing generative AI landscape will depend on its ability to meet this promise. But due to AFM-4.5B, the company made a confident movement.