HTTP 200 verified twice
Finding your next stop…
Loading the latest directory information.
Loading the latest directory information.
AFM-4.5B is a 4.5B-parameter instruction-tuned decoder-only transformer by Arcee.ai, trained on 8 trillion tokens with emphasis on math and code. Apache-2.0 licensed with open weights for cloud-to-edge deployment.
Start here for the decision-making essentials: what AFM 4.5B does, who it is for, how it is accessed, and the first-party sources behind this profile.
See official pricing
Read-only public captures of AFM 4.5B’s homepage and verified pricing page. Screenshots are dated, never live embeds, and open full-screen.
Published specifications and use cases for AFM 4.5B. Repository dates are kept separate from the model's original release date.
Short answers to the questions buyers and builders commonly ask about AFM 4.5B. Each answer cites the shared ledger below, where every source is listed once.
Designed for enterprise-grade deployment across cloud to edge environments
designed for enterprise-grade performance across diverse deployment environments from cloud to edge
4.5B parameters · 5B
Parameters: 4.5B
Text (causal language model, ArceeForCausalLM) · Text Generation
Model Architecture: ArceeForCausalLM
Open weights hosted on Hugging Face at arcee-ai/AFM-4.5B
model_id = "arcee-ai/AFM-4.5B"
Enterprise-grade performance with enhanced capabilities in mathematical reasoning and code
enhanced focus on mathematical reasoning and code generation
Transformer decoder-only (based on Vaswani et al.) with grouped query attention and ReLU^2
The model architecture follows a standard transformer decoder-only design based on Vaswani et al., incorporating several key modifications for enhanced performance and efficiency. Notable architectural features include grouped query attention for improved inference efficiency and ReLU^2 activation functions instead of SwiGLU to enable sparsification while maintaining or exceeding performance benchmarks.
This profile connects the jobs AFM 4.5B is described as handling with its delivery model, access options and the subjects used to match it to related products in this directory.
Each fact points to a recorded source, making it easy to distinguish verified product information from claims that need checking.
HTTP 200 verified twice
arcee-ai/AFM-4.5B · Hugging Face
Model name: AFM-4.5B · Developer: Arcee.ai · Parameter count: 4.5 billion
AFM-4.5B
Arcee.ai
4.5B parameters
Apache-2.0
8 trillion tokens total (6.5T general pretraining + 1.5T midtraining with enhanced focus on mathematical reasoning and code generation)
Transformer decoder-only (based on Vaswani et al.) with grouped query attention and ReLU^2 activation (instead of SwiGLU)
Supervised fine-tuning on instruction datasets followed by reinforcement learning on verifiable rewards and human preference
Designed for enterprise-grade deployment across cloud to edge environments
Open weights hosted on Hugging Face at arcee-ai/AFM-4.5B
A concise view of the jobs, capabilities and integrations described in the recorded product sources.
Documented product formats, platforms and official distribution destinations. Availability can vary by region and plan.
A concise history of software releases and material product changes. Events appear only when a dated source supports what changed.
Showing the newest updates and meaningful milestones. Open an entry for its summary and source.
The public model repository was created on Hugging Face. This repository date may differ from the model's original release date.
View source [1]Facts, answers, structured details, milestones and primary resource links cite this shared ledger. Each external page appears once; release tags from the same GitHub project are grouped under one release history.
These tools share a workflow, capability or audience with AFM 4.5B. They may complement it rather than replace it, and are not presented as integrations or endorsements.
Enterprise-grade performance with enhanced capabilities in mathematical reasoning and code generation
Text (causal language model, ArceeForCausalLM)
AFM-4.5B
arcee-ai
Hugging Face
Text Generation
5B
Base model used for fine-tuning
Hugging Face
AFM-4.5B is a 4.5 billion parameter instruction-tuned model developed by Arcee.ai · Released under the Apache-2.0 license · Transformer decoder-only architecture with grouped query attention
Arcee Foundation Models (AFM)
4.5B
Yes
Transformer decoder-only with grouped query attention and ReLU^2 activation (instead of SwiGLU)
8T tokens total: 6.5T general pretraining + 1.5T midtraining with enhanced focus on mathematical reasoning and code generation
Supervised fine-tuning on high-quality instruction datasets, further refined through reinforcement learning on verifiable rewards and human preference
Enterprise-grade performance across diverse deployment environments from cloud to edge
Hugging Face Transformers, vLLM (>=0.10.1), llama.cpp, Intel OpenVINO, Together AI Playground/API
Text (causal language model)
AFM-4.5B is a 4.5 billion parameter instruction-tuned model developed by Arcee.ai · Parameters: 4.5B · Part of the AFM (Arcee Foundation Models) family
AFM (Arcee Foundation Models)
8 trillion tokens: 6.5T general pretraining plus 1.5T midtraining focused on mathematical reasoning and code generation
Standard transformer decoder-only with grouped query attention and ReLU^2 activation (instead of SwiGLU); model class ArceeForCausalLM
Supervised fine-tuning followed by reinforcement learning on verifiable rewards and for human preference (using modified TorchTitan for pretraining, Axolotl for SFT, and modified Verifiers for RL)
Public open-weights release on Hugging Face Hub (downloadable via transformers, vLLM, llama.cpp, Intel OpenVINO)
Designed for enterprise-grade performance across diverse deployment environments from cloud to edge
Hugging Face transformers library, vLLM (>=0.10.1), llama.cpp, Intel OpenVINO, and Together AI Playground/API
Available via Together AI chat completions API at https://api.together.xyz/v1/chat/completions
