Arcee AI
Top signals
Agent answer
Arcee AI has 247 loaded public signals: 2 hiring, 25 forks, 90 releases or model cards, 105 talking, and 25 repos. Latest signal: arcee-ai/mergekit v0.1.5. Data-business radar is currently scoped to frontier labs, so this category does not expose radar lanes. The standing analysis was generated with deepseek-v4-pro and 91 evidence refs.
has loaded 247 public signals
has hiring signal count 2
has fork signal count 25
has release signal count 90
Thesis
Arcee AI has repositioned from an adapter-and-distillation shop — Arcee Agent was initialized from Qwen2-7B and trained with Spectrum on CrusoeAI compute P4, and it maintains mergekit P3 — into a self-described builder of American, open-weight foundation models, scaling "from a 4.5B dense model to Trinity Large, a 400B MoE model" in six months P6. The wedge is sovereignty, auditability, and inference efficiency for enterprises and institutions that must host and own models rather than consume them as a service P5P27P1. A >$1B Series B (Vista Equity Partners, Cambium Capital, Emergence Capital) and a DOE Genesis-Science-1 trillion-parameter commitment fund a vertically integrated buildout P6W2P27. The distinctive signal is capital efficiency — Fortune reports four models trained for ~$20M W1 — paired with a move down-stack into an open-source agent runtime (nac) and a multi-model API P21P20.
Signal desks
Hiring — Two Greenhouse roles in San Francisco: Technical AI Account Manager (GTM/commercial buildout) E44 and Compute Infrastructure Specialist (training/serving infrastructure) E45. A third-party interview guide for ML and infra engineers stresses merging, distillation, preference tuning, "serving and cost," and "keep[ing] a large training run healthy" on a "small, high-bar team" W3. Hiring evidence is thin — only two job events plus one interview-guide source.
Forks — ansible-slurm (forked from galaxyproject/ansible-slurm), an Ansible role for installing and managing a Slurm HPC cluster, points to self-managed training clusters E17P8. prime-rl (forked from PrimeIntellect-ai/prime-rl), an RL training toolkit, points to a reinforcement-learning infrastructure dependency E57. Only two fork events appear in this pack.
Releases — Trinity family model cards across Nano (6B), Mini (26B), and Large (400B), including base, preview, thinking, and TrueBase variants under "other" license E2E3E4E5E6E7E10E14E16E19; AFM-4.5B variants under Apache-2.0 E20E22E30. Tooling: nac v0.1.0→v0.1.4 plus 0.1.5 RCs and edge builds E33E21E18E48; mergekit v0.1.4/v0.1.5 E59E9; NeMo-RL v1.0.0-rc0/rc1/rc2 P23P22P24; pybubble v0.1.0–v0.4.0 P25E56E50; token.js v0.5.4 P28; vllm rlkit_wheel E58.
Talking — Trinity Large technical report (400B/13B MoE, Muon optimizer, SMEBU stability, DatologyAI synthetic data, 2,048 B300s via Prime Intellect) P1P2; Series B to "build the future of American open models" P6; DOE Genesis-Science-1 trillion-parameter science model P27; NAC harness and Open Models API beta launched the same day P21P20; Bolt on Forge partnership to feed real developer sessions into training/eval P7; science post-training and "AI scientist" framing P26.
Shipping
Shipping concentrates in three artifacts: Hugging Face checkpoints, an open-source agent harness, and a multi-model API. On models, the Trinity family spans Nano (6B), Mini (26B), and Large (400B) with base, preview, thinking, and TrueBase variants; Trinity-Mini leads with 11,154 downloads and 206 likes, while Trinity-Large-Thinking has 2,643 downloads E2E5E3. Earlier AFM-4.5B checkpoints ship Apache-2.0 E20E22E30. The Trinity Large technical report is public as both a blog post and a GitHub repo (124 stars) P1P2E37. nac, a Rust long-running agent harness, is open-sourced under Apache-2.0 and iterates quickly (v0.1.0→v0.1.4 plus release candidates) P21E36E33E21E18. The Open Models API beta adds third-party frontier open models (DeepSeek-V4-Pro/Flash, GLM-5.2, Kimi-K3, Inkling-Small) with per-token pricing and a stated intent to learn task-level model preference to steer Trinity development P20.
Research themes
- Sparse MoE inference efficiency: 400B total / 13B active per token, 256 routed experts + 1 shared with 4 active, 3:1 local-to-global attention, gated attention, QK-normalization, depth-scaled sandwich norm, sigmoid routing P1P2.
- Training stability and optimizers: SMEBU load balancing against router instability/expert collapse, Muon optimizer for larger batches and sample efficiency, zero loss spikes P1P2.
- Synthetic pretraining data at scale: >8T tokens curated by DatologyAI (synthetic web, code, STEM, multilingual), 17T total tokens across three progressively code/math/science-shifted phases P1.
- Reinforcement learning and verifiers: NeMo-RL RCs add verifier environments, a DTensor "v2" backend with 6D parallelism, vLLM-over-HTTP rollouts, and native GRPO tool calling P23; prime-rl fork E57; verifiable RL for science via Trinity-Mini-DrugProt-Think P26.
- Agentic runtimes and "context rot": nac separates transient action context from persistent workstream state, framing harnesses as a new inference runtime P21.
- Merging and distillation: mergekit gains MoE auto-architecture, AFMoE, Qwen3-VL, GLM4-MoE, and IBM Granite support P3; Kimi delta attention distilled into AFM-4.5B E54.
Hiring & scaling
Hiring is concentrated in San Francisco E44E45. Visible roles split between GTM (Technical AI Account Manager E44) and infrastructure (Compute Infrastructure Specialist E45), while the interview guide points to ML/infra needs in merging, distillation, preference tuning, serving cost, and training-run health W3. Scaling evidence: Trinity Large trained on 2,048 NVIDIA B300 GPUs managed by Prime Intellect P1; a forked Slurm cluster Ansible role implies self-managed HPC E17P8; Series B proceeds are earmarked for next-gen Trinity, DOE expansion, and new open-model products P6W2. Fortune's $20M-for-four-models figure frames a capital-efficient scaling thesis W1.
Category implications
- Strategy: Arcee is betting on "American open models" for compliance- and sovereignty-constrained buyers — banks, hospitals, universities, and national labs that need stable, ownable weights P27P5. The explicit contrast is Meta (stopped releasing Llama) and Chinese labs (Western compliance friction) P6P5.
- Infrastructure: self-managed HPC (Slurm fork, B300 fleet, Prime Intellect) plus serving-cost sensitivity in interviews imply heavy compute/ops investment rather than a thin SaaS layer E17P1W3.
- Product: a vertical stack — models → multi-model API (with usage telemetry to steer Trinity) → nac harness → Bolt Forge data flywheel from real development sessions P20P21P7.
- Research: open-weight MoE efficiency, synthetic data, and verifier-based RL are the publishable moat; checkpoints and the tech report are released for external inspection P1P2P23.
- Hiring: a GTM role plus an infra role signals commercialization buildout alongside training — not a research-only lab E44E45.
- GTM: Series B investors (Vista/Cambium/Emergence + Hitachi/Wipro/M12), a DOE national-lab partnership, a Bolt distribution/feedback partnership, and a Trinity Builders program constitute the go-to-market motion P6W2P27P7E46.
Traction highlights
- Trinity-Mini: 11,154 downloads / 206 likes E2; Trinity-Large-Thinking: 2,643 / 190 E3; Trinity-Nano-Preview: 3,354 / 80 E5.
- nac repo: 292 stars E36; trinity-large-tech-report repo: 124–125 stars P2E37.
- Trinity Large launch post drew 231 HN points / 82 comments — the pack's strongest outside-discussion signal E1.
- Series B at >$1B valuation, led by Vista Equity Partners, Cambium Capital, and Emergence Capital P6W2W1.
- Fortune narrative: four models trained for ~$20M W1.