Arcee AI

Arcee AI is an American artificial intelligence company, based in San Francisco, that develops open-weight large language models, chiefly the Trinity family of mixture-of-experts models, and sells API access, hosting and customization services around them.^{[1][3][12][35]} It was founded in 2023 by Mark McQuade, Jacob Solawetz and Brian Benedict.^[2][5] For its first two years Arcee was a post-training and "small language model" vendor that adapted other companies' open models for enterprise customers, and it maintains MergeKit, an open-source toolkit for model merging.^[3][5][6] In 2025 it began pre-training its own models. Between July 2025 and January 2026 it released AFM-4.5B, Trinity Nano, Trinity Mini and the roughly 400-billion-parameter Trinity Large, a lineup the company says was built for about $20 million including salaries, compute, data, infrastructure and operations.^[21][35]

In September 2026 Arcee announced a Series B round led by Vista Equity Partners, Cambium Capital and Emergence Capital, which it said values the company at more than $1 billion. Fortune reported a $1 billion pre-money valuation; Arcee declined to disclose the amount raised, and a source told Fortune it was at least $150 million.^[2][35] The company is also building Genesis-Science-1, an open-weight model for scientific research, with the U.S. Department of Energy under the Genesis Mission.^[29][30]

History

Founding and enterprise post-training (2023-2024)

Arcee came out of stealth in September 2023 with an open-source repository for what it called domain adapted language models (DALM). The code trained a retriever and a generator model jointly, end to end, for retrieval-augmented generation in specialized domains; the launch demos covered patents, PubMed abstracts and SEC filings.^[4] McQuade, the chief executive, had been an early employee at Hugging Face, where he worked on monetization according to Interconnects, and had also worked at Roboflow.^[2][42] TechCrunch reported that the company originally did model customization for large enterprise clients such as SK Telecom. McQuade described the approach: "We would take a Llama model, we would take a Mistral model, we would take a Qwen model that was open source, and we would post-train it to make it better" for a customer's intended use.^[3]

In January 2024 Arcee announced a seed round that brought its total funding to $5.5 million, with investors including Wndrco, Long Journey Ventures and Flybridge.^[5] The next month it merged with mergekit, an open-source model-merging library created by Charles Goddard, who joined Arcee as a senior research engineer.^[5] The paper describing the toolkit, "Arcee's MergeKit: A Toolkit for Merging Large Language Models", lists Goddard and all three co-founders among its authors and appeared in the industry track of EMNLP 2024.^[6] In July 2024 Arcee raised a $24 million Series A led by Emergence Capital, with seed investors Long Journey Ventures, Scott Banister, Flybridge and Centre Street Partners and new investor Arcadia Capital. It launched Arcee Cloud, a hosted service for training, merging and deploying custom models, and hired Julien Simon from Hugging Face as chief evangelist.^[7][8]

Through 2024 and early 2025 Arcee released post-trained and distilled models built on other labs' open bases. Arcee-SuperNova (September 2024) was a 70B model that merged a distillation of Llama 3.1 405B Instruct into Llama 3.1 70B Instruct with separately instruction-tuned and preference-tuned variants.^[9] In November 2024 Arcee signed a strategic collaboration agreement with Amazon Web Services to offer its models to AWS customers, including through Amazon SageMaker; the announcement named Guild Education as a customer.^[11] The same month it reported contributing 8,200 H100 GPU-hours and post-training work (supervised fine-tuning, DPO, merging and logit distillation) to Prime Intellect's decentralized INTELLECT-1 model.^[12] In January 2025 it published Virtuoso-Lite (10B, built on TII's Falcon architecture) and Virtuoso-Medium-v2 (32B), both distilled from DeepSeek V3.^[10]

Decision to pre-train (2025)

In June 2025 Arcee introduced the Arcee Foundation Model (AFM) family with AFM-4.5B, a dense 4.5-billion-parameter model trained on data curated by DatologyAI.^[13] Around the same time it open-sourced five earlier models, three of which had been available only through its SaaS platform (Arcee-SuperNova-v1, Virtuoso-Large and Caller) and two research models (GLM-4-32B-Base-32K and Homunculus), describing the release as part of a shift in focus to AFM.^[14] The weights of AFM-4.5B and AFM-4.5B-Base were published on Hugging Face on July 29, 2025. The base model saw 8 trillion tokens, 6.5 trillion of general pretraining data followed by 1.5 trillion of mid-training data weighted toward mathematics and code, and the architecture used grouped-query attention and ReLU² activations.^[15] At launch the model used a custom Arcee Model License that allowed free commercial use for companies with less than $1.75 million in annual revenue, in place of a previously planned non-commercial CC-BY-NC license.^[15] Arcee later moved AFM-4.5B to the Apache License 2.0, which its model card now lists.^[16][19]

On July 30, 2025 Arcee announced a strategic funding round led by Prosperity7 Ventures, the venture fund of Aramco Ventures, and M12, Microsoft's venture fund, with participation from Hitachi Ventures, Wipro, JC2 Ventures, Samsung Next and Guidepoint chief executive Albert Sebag. It did not disclose the amount. The announcement still described Arcee as an enterprise platform built on compact models and said the next AFM release would be a sparse model of about 120 to 140 billion parameters with 20 to 30 billion active.^[17]

The plan grew larger. Fortune later reported that McQuade saw an opening after Meta backed away from open-weight models, and that the company "had $30 million in the bank" when he decided to spend 65% to 70% of its capital on building models from scratch.^[2] In the December 2025 "Trinity Manifesto", chief technology officer Lucas Atkins gave three reasons for owning pretraining: diminishing returns from post-training on some high-stakes workloads, enterprise buyers asking where a base model came from and what data and licenses governed it, and a belief that future applications would keep training on live usage, which requires control of the weights.^[20]

MergeKit's license changed during this period. Version 0.1, released in February 2025, moved the library from the GNU Lesser General Public License to the Business Source License (BSL).^[18] On October 31, 2025 Arcee returned it to LGPL v3, saying the custom license had created legal review friction for developers and was out of step with its permissively licensed models.^[19]

Trinity (December 2025 to April 2026)

On December 1, 2025 Arcee released Trinity Nano Preview, a 6B-parameter MoE with about 1B active parameters, and Trinity Mini, a 26B MoE with 3B active, under Apache 2.0. Both were trained on 10 trillion tokens on 512 H200 GPUs. Mini was offered through Arcee's API and OpenRouter; Nano Preview, which Arcee called an experimental, "personality-forward" chat model that "may be unstable in edge cases", was download-only.^[20] At the time Trinity Large was training on 2,048 B300 GPUs, with compute partner Prime Intellect operating the cluster.^[20][23]

Trinity Large was released on January 27, 2026 as three checkpoints: Trinity-Large-Preview, a lightly post-trained instruct model; Trinity-Large-Base, the final pretraining checkpoint after 17 trillion tokens; and Trinity-Large-TrueBase, a 10-trillion-token checkpoint with no instruction data or learning-rate annealing, intended for researchers who want to study raw pretraining.^[21] Pretraining took 33 days on 2,048 NVIDIA B300 GPUs, which Arcee said was, as far as it could tell, the largest publicly stated pretraining run on those machines.^[21] The preview was free on OpenRouter during the preview period and was integrated at launch with the coding agents Kilo Code, Cline and OpenCode.^[21] TechCrunch, which described Arcee as a 30-person startup, noted that the model handled text only; Atkins said a vision model was in development and speech-to-text was on the roadmap.^[3] A technical report was posted to arXiv in February 2026.^[22]

Trinity-Large-Thinking, a reasoning version that produces a chain of thought before answering, followed on April 1, 2026 on Arcee's API and as Apache 2.0 weights on Hugging Face. Arcee said its post-training ran on 1,152 H100 GPUs and that serving used NVIDIA Dynamo, Blackwell Ultra GPUs and vLLM.^[24] The company reported that Trinity-Large-Preview had served 3.37 trillion tokens on OpenRouter in its first two months and was the most used open model in the U.S. and fourth globally in OpenRouter's OpenClaw collection. It also said Trinity-Large-Thinking ranked second on PinchBench, an agent benchmark from Kilo, behind Claude Opus 4.6, at $0.90 per million output tokens on its API.^[24] McQuade told TechCrunch it was the most capable open-weight model "ever released by a non-Chinese company"; TechCrunch then described Arcee as a 26-person company.^[26]

On May 29, 2026 Arcee moved the whole Trinity family, including quantized variants and already-released checkpoints, from Apache 2.0 to OpenMDW-1.1, the Linux Foundation's license for model distributions, saying it covered weights, code, configuration and documentation under one instrument without adding field-of-use restrictions.^[27]

Expansion and Series B (2026)

On June 9, 2026 Hugging Face announced what it called a multi-million dollar commercial collaboration under which the Hugging Face Hub became the exclusive home for all of Arcee's models, datasets and agent traces, public and private, with private artifacts stored in Hugging Face's Buckets product in place of Amazon S3. Hugging Face said the arcee-ai organization had published 204 models and 63 datasets.^[28]

On July 22, 2026 Arcee and the Department of Energy announced Genesis-Science-1 (GS1), an open-weight model for scientific research developed with the national laboratories. Arcee said it had secured the compute and would handle training, post-training and the surrounding software, while DOE scientists would choose problems, supply data and environments, and evaluate results. It described GS1 as a "trillion-parameter-class" language model built on its next generation of Trinity models, to be released openly later in 2026 with weights, a technical report and demonstrations.^[29] On August 7, 2026 the DOE launched a Genesis Open Models Initiative, saying the first model in the class "will be Genesis-Science-1, developed in partnership with Arcee", and invited outside contributions of models, pretraining data and fine-tuning work.^[30]

On August 13, 2026 Arcee open-sourced nac, an agent harness for long-running tasks, and expanded its Open Models API beta beyond Trinity to models from other labs, including DeepSeek V4-Pro, GLM-5.2, Kimi K3 and Inkling-Small from Thinking Machines Lab.^[32][33] The API beta listed Trinity-Large-Thinking at $0.25 per million input tokens and $0.80 per million output tokens.^[33] From September 14 to October 14, 2026 it ran Forge, a research preview with Bolt.new in which opted-in Bolt Pro users received more usage of selected open models and their sessions could be used to train and evaluate future Arcee open-weight models.^[34]

Arcee announced its Series B on September 16, 2026, with participation from AI10 Ventures, Hitachi, IAG, M12, P7 and Wipro alongside the three lead investors. It said the money would fund the next generation of Trinity, "already in training", expand its work with the Department of Energy and build products around open models.^[35] Fortune reported that Arcee would also work extensively with Vista's portfolio companies. McQuade told Fortune: "We're not chasing [Poolside's] Laguna. We're chasing [Beijing-based Z.ai's] GLM Flash."^[2]

Models

The Hugging Face model cards give Trinity Large's size as about 398 billion parameters.^[25] Arcee also publishes quantized versions of the Trinity checkpoints (FP8, NVFP4, W4A16 and GGUF formats) on Hugging Face.^[27][43]

Trinity architecture and training

Arcee calls the Trinity architecture "afmoe". The December 2025 manifesto and the technical report describe a combination of published techniques: grouped-query attention, RMSNorm applied to queries and keys (QK-norm), gated attention after scaled dot-product attention, and a 3:1 interleave of local sliding-window attention layers with RoPE and global attention layers with no positional embedding. Each sublayer uses a depth-scaled "sandwich" normalization. The MoE layers follow the DeepSeekMoE design of fine-grained routed experts plus one always-active shared expert, with sigmoid routing and the auxiliary-loss-free load balancing introduced in DeepSeek-V3. The models were trained with the Muon optimizer on a modified version of TorchTitan in bf16 precision.^[20][22]

Trinity Large is unusually sparse. It routes each token to 4 of 256 experts, a routing fraction of 1.56%, compared with 3.13% for DeepSeek-V3 (8 of 256) and 0.78% for Llama 4 Maverick (1 of 128), according to Arcee.^[21] The company originally targeted 420 billion parameters but raised the number of dense layers from three to six to keep routing stable at that sparsity.^[21] The technical report introduces two stability methods: Soft-clamped Momentum Expert Bias Updates (SMEBU), which replaces fixed-size router bias steps with bounded, momentum-smoothed updates, and a Random Sequential Document Buffer (RSDB) data loader that Arcee says cut batch heterogeneity by 4.23 times and step-to-step loss variance by 2.4 times in the final phase. The report says all three Trinity models finished pretraining with zero loss spikes.^[22][23]

DatologyAI curated Trinity Large's pretraining data: 17 trillion tokens in three phases of 10, 4 and 3 trillion tokens, including more than 8 trillion tokens of synthetic data and data in 14 languages other than English.^[21] Arcee's December 2025 post had described a planned 20-trillion-token dataset, and DatologyAI's case study describes the corpus it produced as 20 trillion tokens (8 trillion synthetic and 12 trillion web); the published pretraining run used 17 trillion of them.^[20][22][37] Arcee extended the context window to 256K tokens by adjusting only the global attention layers. It reports multi-key needle-in-a-haystack scores of 0.994 at 256K, 0.976 at 512K (a length it never trained on) and 0.42 at 1 million tokens.^[23] NVIDIA's case study adds that Arcee used NVIDIA NeMo RL for reinforcement-learning post-training of Nano and Mini and NeMo Data Designer for data preparation.^[36]

Reported benchmark results

All figures below are Arcee's own evaluations.

Source: Arcee's January 2026 launch post.^[21]

Source: Trinity-Large-Thinking model card.^[25] On this table Trinity-Large-Thinking has the top score only on τ²-bench Airline, and it trails Claude Opus 4.6 on every benchmark except the two τ²-bench tasks.

Open-source tools and products

• MergeKit is a library for combining the weights of several fine-tuned models without further training, supporting methods such as SLERP, task arithmetic, TIES-Merging and DARE. Its repository dates to August 2023, and Arcee has maintained it since the February 2024 merger with Goddard's project.^[5][6][38] Arcee says IBM Research used it for merging ablations during development of IBM Granite 4.0.^[39]
• DistillKit and EvolKit are Arcee libraries for logit-based knowledge distillation and evolutionary synthetic instruction generation; both were used to build Arcee-SuperNova.^[9]
• Spectrum is a 2024 method, co-authored by Atkins, that speeds up post-training by choosing which layer modules to train from their signal-to-noise ratio and freezing the rest; Arcee lists "spectrum training" among the techniques used for AFM-4.5B.^[17][40]
• nac is an Apache-2.0 agent harness released in August 2026. A central orchestrator only plans and dispatches work to "threads", each run by a fresh worker that returns a structured summary ("episode") of what it did; Arcee says the design adapts the thread-and-episode architecture from Random Labs' Slate report.^[32]
• Hosted products have included Arcee Cloud (2024), the Arcee Conductor model-routing platform, the Orchestra environment for building agent workflows, and the Arcee API and chat site that serve the Trinity models.^{[7][14][17][20]}

Funding

In January 2026 McQuade told TechCrunch the company had raised about $50 million in total.^[3] For the Series B, Arcee said the round values it at "more than $1 billion", and Fortune reported a $1 billion pre-money valuation; neither gave a post-money figure.^[2][35]

Partnerships and customers

Arcee's model work depends on a few outside partners. DatologyAI has curated the pretraining data for every Arcee model since AFM-4.5B, and Atkins described the partnership to Interconnects as "almost an extension of our own research team."^[15][37][42] Prime Intellect supplied compute for synthetic data generation and the Trinity Large run.^[20][23] NVIDIA published a case study on the Trinity Large training and inference stack, and said Arcee planned to build an NVIDIA-powered "AI factory".^[36] Other announced relationships include the 2024 AWS agreement, a July 2025 joint venture with Dubai's Seed Group to sell Arcee's enterprise products in the Middle East and North Africa, the 2026 Hugging Face storage agreement, and the Department of Energy's GS1 project.^{[11][28][29][41]} Named enterprise customers have included SK Telecom and Guild Education.^[3][11]

Views on open models and China

Arcee frames its work as supplying American open-weight models at a time when Chinese labs dominate open releases. The Trinity Manifesto said Qwen, DeepSeek and others "now define a lot of what 'state of the art open MoE' looks like", while in the United States "most of the action has centered on polishing other people's checkpoints."^[20] The Series B announcement called Trinity Large the first highly capable, permissively licensed model "developed end to end in the United States" since, in Arcee's account, Meta stopped releasing Llama models in early 2025.^[35]

The company has argued against treating Chinese models as inherently unsafe. In a July 2026 TechCrunch interview Atkins said that once a model runs in a customer's own environment "there is really not any way for an Arcee, or an Alibaba" to access it, and that "instead of the conversation being about how to ban Chinese models, it should be about how do we foster a good, open ecosystem here in the U.S." TechCrunch noted that if any startup would benefit from a ban on Chinese models, it would be Arcee.^[31] Arcee's GS1 announcement likewise expressed "real admiration" for DeepSeek, Qwen, Kimi, MiniMax and GLM while arguing that institutions with sensitive work also weigh who trained a model and under which country's laws.^[29]

Reception

Coverage has focused on the company's small budget and headcount. TechCrunch's Julie Bort wrote in April 2026 that she "can't help rooting" for Arcee, while noting that its models "are not outperforming the closed source models from the big labs like Anthropic or OpenAI."^[26] Earlier she wrote that Trinity Large was "not a true SOTA competitor yet because it currently supports only text."^[3] Nathan Lambert of Interconnects called Arcee "the startup I've found to be taking the most real approach to monetizing their open models."^[42] Fortune highlighted the $20 million training cost against conventional assumptions that frontier models cost billions to train, and quoted McQuade: "We can comfortably say we're the most efficient lab in the world based on what we've done."^[2] Arcee's own published results show Trinity-Large-Thinking behind Claude Opus 4.6 on most agentic and coding benchmarks, including a gap of about 12 points on SWE-bench Verified, and behind GLM-5, MiniMax-M2.7 and Kimi-K2.5 on IFBench.^[25]

Leadership

See also

• List of neolabs
• Neolab
• Model merging
• DatologyAI
• Prime Intellect
• Open-source AI
• Small language model

References

1. ^Arcee AI. "About Arcee AI." Accessed October 27, 2026. arcee.ai/about
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Allie Garfinkle. "Arcee AI trained four models for $20 million. Now, it's worth $1 billion." Fortune, September 16, 2026. fortune.com/...-20-million-now-its-worth-1-billion
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Julie Bort. "Tiny startup Arcee AI built a 400B-parameter open source LLM from scratch to best Meta's Llama." TechCrunch, January 28, 2026. techcrunch.com/...from-scratch-to-best-metas-llama
4. ^1 ^2Jacob Solawetz. "Arcee Exits Stealth with Open Core LLMs." Arcee AI blog, September 19, 2023 (archived). web.archive.org/...s-stealth-with-open-core-llms-2
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Arcee AI. "Arcee celebrates seed round & merger with mergekit." Arcee AI blog, 2024 (archived). web.archive.org/...seedroundandmergekitmerger
6. ^1 ^2 ^3 ^4 ^5 ^6Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi, Luke Meyers, Vladimir Karpukhin, Brian Benedict, Mark McQuade, Jacob Solawetz. "Arcee's MergeKit: A Toolkit for Merging Large Language Models." Proceedings of EMNLP 2024: Industry Track, pages 477-485, November 2024. aclanthology.org/2024.emnlp-industry.36 (arXiv:2403.13257)
7. ^1 ^2 ^3Arcee AI. "Our Series A, Julien Simon Joins the Team, & Arcee Cloud Goes Live." Arcee AI blog, July 2024 (archived). web.archive.org/...-the-team-arcee-cloud-goes-live
8. ^1 ^2Gunderson Dettmer. "Gunderson Dettmer Advises Arcee AI in its $24M Series A Financing." July 16, 2024. gunder.com/...nounces-dollar24m-series-a-financing
9. ^1 ^2 ^3Lucas Atkins and Fernando Fernandes. "Meet Arcee-SuperNova: Our Flagship 70B Model, Alternative to OpenAI." Arcee AI blog, September 10, 2024. arcee.ai/...agship-70b-model-alternative-to-openai
10. ^Arcee AI. "Virtuoso-Lite & Virtuoso-Medium-v2: Distilling DeepSeek-V3 into 10B & 32B Small Language Models (SLMs)." Arcee AI blog, January 28, 2025 (archived). web.archive.org/...-32b-small-language-models-slms
11. ^1 ^2 ^3Arcee AI. "Arcee AI Signs Strategic Collaboration Agreement with AWS to Accelerate the Deployment of Smaller, Specialized Language Models." Arcee AI blog (Business Wire release), November 24, 2024. arcee.ai/...of-smaller-specialized-language-models
12. ^1 ^2Lucas Atkins, Charles Goddard, Maziyar Panahi, Fernando Fernandes. "Arcee AI's Role in the INTELLECT-1 Open Source Revolution." Arcee AI blog, November 29, 2024. arcee.ai/...arcee-and-the-intellect-1-2
13. ^1 ^2Arcee AI. "Announcing the Arcee Foundation Model family." Arcee AI blog, June 2025 (archived June 30, 2025). web.archive.org/...e-arcee-foundation-model-family
14. ^1 ^2Arcee AI. "Releasing five new open-weights models: SuperNova 70B, Virtuoso-Large 72B, Caller 32B, GLM-4-32B-Base-32K, and Homunculus 12B." Arcee AI blog, 2025 (archived July 1, 2025). web.archive.org/...ng-five-new-open-weights-models
15. ^1 ^2 ^3 ^4Lucas Atkins. "Announcing the Official Launch of AFM-4.5B." Arcee AI blog, July 29, 2025. arcee.ai/...uncing-the-official-launch-of-afm-4-5b
16. ^Arcee AI. "AFM-4.5B" model card. Hugging Face. Accessed October 27, 2026. huggingface.co/...AFM-4.5B
17. ^1 ^2 ^3 ^4Mark McQuade. "Arcee AI Secures Strategic Investment to Accelerate Enterprise-Grade AI Platform Built on AFM Foundation Models." Arcee AI blog, July 30, 2025. arcee.ai/...-announces-new-strategic-funding-round
18. ^arcee-ai/mergekit. "MergeKit v0.1: Universal model support, Arcee Fusion, multi-GPU, and BSL license" (commit to LICENSE). GitHub, February 5, 2025. github.com/...0b6f2c67270ec97f514740d8ff214d2a46c6
19. ^1 ^2Arcee AI. "Mergekit Returns To Its Roots." Arcee AI blog, October 31, 2025 (archived). web.archive.org/...mergekit-returns-to-its-roots
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Lucas Atkins. "The Trinity Manifesto: Introducing Trinity Mini." Arcee AI blog, December 1, 2025. arcee.ai/...the-trinity-manifesto
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Lucas Atkins. "Trinity Large." Arcee AI blog, January 27, 2026. arcee.ai/...trinity-large
22. ^1 ^2 ^3 ^4 ^5Varun Singh, Lucas Krauss, Sami Jaghouar, et al. "Arcee Trinity Large Technical Report." arXiv:2602.17004, February 19, 2026. arxiv.org/...2602.17004
23. ^1 ^2 ^3 ^4Lucas Atkins, Varun Singh, Charles Goddard, Mark McQuade. "Trinity Large Tech Report: 400B Open MoE Architecture Explained." Arcee AI blog, February 20, 2026. arcee.ai/...trinity-large-tech-report
24. ^1 ^2 ^3Lucas Atkins. "Trinity-Large-Thinking: Scaling an Open Source Frontier Agent." Arcee AI blog, April 1, 2026. arcee.ai/...trinity-large-thinking
25. ^1 ^2 ^3 ^4Arcee AI. "Trinity-Large-Thinking" model card. Hugging Face. Accessed October 27, 2026. huggingface.co/...Trinity-Large-Thinking
26. ^1 ^2Julie Bort. "I can't help rooting for tiny open source AI model maker Arcee." TechCrunch, April 7, 2026. techcrunch.com/...open-source-ai-model-maker-arcee
27. ^1 ^2Lucas Atkins. "Trinity is moving to OpenMDW-1.1." Arcee AI blog, May 29, 2026. arcee.ai/...trinity-is-moving-to-openmdw-1-1
28. ^1 ^2Hugging Face. "Arcee Becomes the First Major American AI Lab to Replace AWS S3 with Hugging Face Private Storage, in a Multi-Million Dollar Commercial Partnership." Hugging Face blog, June 9, 2026. huggingface.co/...arcee-hf
29. ^1 ^2 ^3 ^4 ^5Lucas Atkins, Mark McQuade and the Arcee AI Team. "Genesis-Science-1." Arcee AI blog, July 22, 2026. arcee.ai/...genesis-science-1
30. ^1 ^2U.S. Department of Energy. "U.S. Department of Energy Launches the Genesis Open Models Initiative." August 7, 2026. energy.gov/...nches-genesis-open-models-initiative
31. ^Julie Bort. "Arcee, a US open source AI lab, says Chinese models are not inherently dangerous." TechCrunch, July 22, 2026. techcrunch.com/...els-are-not-inherently-dangerous
32. ^1 ^2The Arcee AI Team, Brett Larsen, Lucas Atkins, Varun Singh. "Introducing NAC, an Open-Source Harness for Long-Running Agent Work." Arcee AI blog, August 13, 2026. arcee.ai/...nac
33. ^1 ^2The Arcee AI Team. "Introducing the Arcee Open Models API Beta." Arcee AI blog, August 13, 2026. arcee.ai/...open-models-api-beta
34. ^The Arcee AI Team. "Arcee Partners with Bolt on Forge." Arcee AI blog, September 14, 2026. arcee.ai/...arcee-partners-with-bolt-on-forge
35. ^1 ^2 ^3 ^4 ^5 ^6 ^7Lucas Atkins, Mark McQuade and the Arcee AI Team. "Arcee AI Raises Series B to Build the Future of American Open Models." Arcee AI blog, September 16, 2026. arcee.ai/...ild-the-future-of-american-open-models
36. ^1 ^2NVIDIA. "Arcee Trains Frontier Open Models for Long-Running Agents, Powered by NVIDIA Blackwell" (case study). Accessed October 27, 2026. nvidia.com/...arcee-ai
37. ^1 ^2DatologyAI. "Faster, Better, Smaller: Arcee AI works with DatologyAI to release their foundation model." DatologyAI blog, April 10, 2026. datologyai.com/...arcee-case-study
38. ^arcee-ai/mergekit. GitHub repository. Accessed October 27, 2026. github.com/...mergekit
39. ^Arcee AI. "IBM Research Uses Arcee MergeKit in Granite 4.0 Model Development." Arcee AI blog, October 2025 (archived). web.archive.org/...-in-granite-4-model-development
40. ^Eric Hartford, Lucas Atkins, Fernando Fernandes Neto, David Golchinfar. "Spectrum: Targeted Training on Signal to Noise Ratio." arXiv:2406.06623, June 2024. arxiv.org/...2406.06623
41. ^Arcee AI. "Seed Group teams up with Arcee AI to revolutionise enterprise AI innovation in Dubai and the wider MENA region." Arcee AI blog, July 22, 2025 (archived). web.archive.org/...dubai-and-the-wider-mena-region
42. ^1 ^2 ^3 ^4Nathan Lambert. "Arcee AI goes all-in on open models built in the U.S." Interconnects, January 27, 2026. interconnects.ai/...-ai-goes-all-in-on-open-models
43. ^Arcee AI (arcee-ai) organization page. Hugging Face. Accessed October 27, 2026. huggingface.co/arcee-ai