Arcee.ai

Loka, AWS and Arcee post-trained Trinity Mini across two reinforcement-learning environments for biomedical research: one rewarding tool choice and evidence-seeking, one rewarding Gene Ontology inference in strict JSON. 21 controlled runs, with prompt-search gains reported apart from held-out result

Arcee's Trinity-Large-Thinking is live on its own API, OpenRouter and Hugging Face under Apache 2.0. Two months of SFT and RL work followed Preview, aimed at multi-turn tool use, context coherence across turns, instruction following under constraint, and stable long-running behavior.

Charles Goddard details how AFM-4.5B's context window went from 4k to 64k through heavy experimentation with model merging, distillation and repeated model soups, rather than a single dedicated long-context training run.

Arcee researchers Charles Goddard and Fernando Fernandes Neto describe tokenizer transplantation: using orthogonal matching pursuit to express one model's embeddings in another's vocabulary, so models with different tokenizers can be merged, distilled or paired for speculative decoding without retra

Arcee released Trinity Large as three variants from one training run: a chat-ready Preview, a full 17T-token Base, and a 10T-token TrueBase with no instruct data or learning rate anneals. The MoE routes 4 of 256 experts, trained on 2,048 B300 GPUs in 33 days using z-loss and momentum load balancing.

Arcee-Nova is a 72B open model built by merging Qwen2-72B-Instruct with a custom-tuned generalist and layering RLHF on top, released with GGUF quantisations and scored on the OpenLLM Leaderboard 2.0 evaluation stack.

Arcee's 400B sparse MoE, trained on 2048 B300 GPUs in about 30 days, released as Preview, Base and a 10T-token TrueBase checkpoint. Covers the 4-of-256 routing choice, extra dense layers for routing stability, momentum-based expert balancing and z-loss.

Arcee is handing out inference credits on its API to open-source projects, research groups, prototypes and tooling builders, sized from under 50M to over 1B tokens per grant. Applications are reviewed on a best-effort basis against available capacity.

Arcee open-sources nac, an agent harness for tasks that run for hours or days. The design splits ephemeral action context from durable workstream state: an orchestrator that can only dispatch threads, which report back structured episodes instead of raw transcripts.

Every Trinity model, including previously shipped weights and quantizations, moves to OpenMDW-1.1, a Linux Foundation license that covers weights, configs, evaluation materials, and data as a single distribution instead of stretching a code license over them. NVIDIA plans to follow.

Arcee-VyLinh is a 3B Vietnamese model built from Qwen2.5-3B via EvolKit-generated hard questions answered by a 72B teacher, SFT, a merge with Qwen2.5-3B-Instruct, and DPO refinement, shipped with a quantised build for consumer hardware.

Arcee open-sourced DistillKit, the toolkit behind Virtuoso and SuperNova Medius. It runs online and offline distillation from a single YAML config, compresses stored teacher logits to roughly 300 bytes per token, and ships captured teacher data from Qwen3-235B and DeepSeek V3/R1.

Arcee introduced its Foundation Model family, starting with AFM-4.5B, a 4.5 billion parameter model targeted at running on phones, edge hardware or cloud. It is testable in Arcee's playground and on Together AI, with training pipeline and benchmark detail in a companion technical post.

Arcee released open weights for five models: the production systems SuperNova, Virtuoso-Large and Caller that had been serving customer workloads on its SaaS, plus research models GLM-4-32B-Base-32K and Homunculus, as the company shifts focus to its AFM foundation line.

Trinity-Large-Thinking is free on OpenRouter until May 23, positioned for multi-turn tool use and long-horizon routing. Arcee notes the model is tuned to carry reasoning across calls, which OpenRouter documents as interleaved thinking.

A walkthrough of serving AFM-4.5B on Intel Xeon 6 using OpenVINO and Hugging Face Optimum Intel: quantize and convert in a single CLI command, then serve through OpenVINO Model Server behind an OpenAI-compatible endpoint, benchmarked on an r8i.8xlarge instance.

Arcee took AFM-4.5B out of preview and published weights for both the instruction-tuned model and AFM-4.5B-Base on Hugging Face, moving its first foundation model from a hosted preview to something teams can download, run across cloud or edge, and fine-tune themselves.

Arcee released five models as open weights: SuperNova-v1 70B, Virtuoso-Large 72B, the tool-calling Caller 32B, GLM-4-32B-Base-32K and Homunculus 12B. Three had been production-only on its SaaS, with licenses ranging from Llama 3 Community to Apache-2.0.

Arcee is building Genesis-Science-1 with the US Department of Energy: an open-weight model paired with a governed research harness meant to run scientific computing workflows while keeping a reproducible record of its work. A contributor program opens to labs and universities.

Arcee is switching off its hosted Trinity-Large-Preview endpoint on April 22, 2026, redirecting constrained serving capacity to paid Trinity-Large-Thinking users. The team says it would work with any outside provider willing to host Preview instead.

Arcee opens a Model-Tools repository of LLMOps utilities: a GUI over the LM Eval Harness, a model difference examiner that reports shape and config mismatches between checkpoints, and an interactive base-model-plus-adapter merger.

Intel's OpenVINO 2025.3 release adds full support for Arcee's AFM-4.5B, and quantized 16-bit, 8-bit and 4-bit builds are published on Hugging Face. That covers Intel CPUs and GPUs from edge devices through cloud instances.

Arcee has moved its Trinity open-weight models onto the OpenMDW-1.1 license, changing the terms attached to the released weights.

Spectrum trains only the layers with a high signal-to-noise ratio and freezes the rest. Arcee reports up to 42 percent faster training with quality held, lower memory use allowing larger batches, and less catastrophic forgetting because untouched layers keep prior knowledge.