Nous Research

Thesis

Nous Research is in the midst of a rapid productization pivot. Hermes Agent — a 190k-star open-source AI agent platform — is receiving weekly releases with ~1,475 commits per cycle and 245 community contributors P1P24. Simultaneously, the org is shipping frontier models (Hermes-4 family through 405B scale), building RL training infrastructure (Atropos at 1,273 stars), and advancing distributed training research (DisTrO/Psyche Network). A cluster of May 2026 NVIDIA infrastructure forks (Megatron-LM, Megatron-Bridge, NeMo RL, NeMo Gym, Nemotron) alongside a General Counsel hire and Forward Deployed Engineer role signal a lab moving from research collective toward platform company with enterprise posture E19E20E22E29E30E59E60.

Signal desks

Hiring

• UI/UX Designer — open role implies consumer-facing product buildout for Hermes Desktop and agent interfaces E19.
• General Counsel — first in-house legal hire, signals platform maturity, regulatory posture, and/or commercial contracting needs E20.
• Full Stack Engineer — product engineering for web/desktop surfaces around Hermes Agent E21.
• Forward Deployed Engineer — enterprise/partner deployment role, suggests organizational customers onboarding onto Hermes or Forge infrastructure E22.
• Research Scientist — continued fundamental research capacity, likely spanning RL (Atropos), long-context attention, and reasoning E23.
• Machine Learning Engineer — training and fine-tuning pipeline scaling for the Hermes-4 family, MoE models, and Atropos RL training E24.

Forks

• NVIDIA training/infra cluster — Nous forked NVIDIA's Megatron-LM, Megatron-Bridge, Automodel, NeMo RL, NeMo Gym, and Nemotron in late April–May 2026, suggesting large-scale training infrastructure buildout or inspection for Hermes model family and Atropos RL workloads E28E29E30E46E59E60.
• NVIDIA desktop/shell tools — OpenShell, OpenShell-Community, and NemoClaw forks (April 2026) point to Hermes Desktop terminal and sandbox capabilities under active development E47E51E52.
• Microsoft agent-governance-toolkit — fork (May 2026, 31 stars) signals interest in enterprise agent governance, compliance, and controlled-agent deployment patterns E37.
• vercel-labs/wterm and vadimdemedes/ink — terminal-UI tools forked in April 2026, consistent with Hermes Desktop surface development E50E58.

Releases

• Hermes Agent v0.17.0 (v2026.6.19) — "The Reach Release": iMessage channel via Photon, Raft agent network gateway, upgraded subagents, image generation editing, Cursor Composer via xAI Grok, dashboard profile builder, major Skills Hub rework, memory tool overhaul, 300+ issues closed P1E16.
• Hermes Agent v0.16.0 (v2026.6.5) — "The Surface Release": Hermes Desktop public preview on Windows/macOS/Linux; NVIDIA/skills added as built-in trusted skills tap alongside OpenAI, Anthropic, HuggingFace W2E18.
• Hermes Agent weekly cadence — versions v2026.5.29.2, v2026.5.29, v2026.5.28, v2026.5.16, v2026.5.7, v2026.4.30, v2026.4.23, v2026.4.16, v2026.4.13 all shipped between mid-April and early June 2026 E25E26E27E35E41E44E48E53E57.
• Model: Hermes-4-405B (Aug 2025) — 405B parameter hybrid-reasoning chat model, llama3 license, 286 downloads E15.
• Model: Hermes-4-70B (Aug 2025) — 70B frontier chat model, 4,062 downloads E8.
• Model: Hermes-4-14B (Aug 2025) — 14B model, Apache-2.0 license, 31,365 downloads E11.
• Model: Hermes-4.3-36B (Nov 2025) — 36B, Apache-2.0, 25,308 downloads E6.
• Model: NousCoder-14B (Dec 2025) — code-focused model, Apache-2.0, 212 likes E7.
• Model: nomos-1 (Dec 2025) — 30.5B reasoning model for mathematical proof-writing, Apache-2.0 E12P25.
• Model: Minos-v1 (Apr 2025) — 396M text-classification model (likely a reward/judge model), Apache-2.0 E10.
• Model: moe-10b-a1b-8k-wsd-lr3e4-1t (Mar 2026) — experimental MoE architecture, 10B total / ~1B active parameters E45.
• DeepHermes Atropos specialist models — ToolCalling-Specialist, Financial-Fundamentals-Prediction-Specialist, and AscensionMaze-RLAIF variants, all 8B Llama-3-based, pipeline tagged reinforcement-learning E39E43E54.
• Atropos RL Framework (v2025.4.29+) — 1,273 stars, 367 forks, LLM RL gym for trajectory collection and evaluation, active as of June 2026 P21E33.
• tinker-atropos — integration layer connecting Atropos with Thinking Machines Tinker API for managed RL training P26.
• Hermes Agent self-evolution (Mar 2026) — 4,364-star repo: DSPy + GEPA framework for optimizing agent skills, prompts, and code E36.
• Hermes paperclip-adapter (Mar 2026) — 1,650-star TypeScript adapter to run Hermes as a managed employee in Paperclip companies E55.
• hermes-compression-eval (May 2026) — offline probe-based evaluation harness for Hermes Agent's ContextCompressor, adapted from Factory's compression evaluation methodology E34.
• kanban-video-pipeline (May 2026) — 43-star new repo, likely a media generation pipeline E42.

Talking

• Lighthouse Attention — Nous proposed and open-sourced a selection-based hierarchical attention mechanism delivering 1.4–1.7× pretraining speedup at 98K context and ~17× faster than standard attention at 512K on a single B200. Covered by MarkTechPost and Newsference in mid-May 2026 W3W4.
• Hermes Desktop launch — The Decoder covered the Hermes Desktop public preview (MIT license, Windows/macOS/Linux), calling it "an open-source AI agent for every platform," highlighting sub-agents, sandboxed execution (local/Docker/SSH/Singularity/Modal), and access to 300+ models via Nous Portal W6.
• Nous Portal integration — teknium1 authored a dedicated Portal integration page, positioning Hermes-4-70B and Hermes-4-405B as "frontier hybrid-reasoning chat models" available at heavily discounted rates, alongside Google Gemini, Anthropic Claude, and other providers through OpenRouter-style aggregation W1.
• GitHub org overview — as of late May 2026, NousResearch's top repos by stars: hermes-agent (200k+), hermes-agent-self-evolution (4.3k), hermes-paperclip-adapter (1.6k), Atropos (1.3k+), plus autonovel experimental writing pipeline W5.

Shipping

The pace of shipping is aggressive and multi-vector. Hermes Agent is on a weekly release cadence — at least 8 versioned releases between April 13 and June 19, 2026, with v0.17.0 alone representing ~1,475 commits, ~800 merged PRs, and 245 community contributors P1E16E25E26E27E35E41E44E48E53E57. The agent has expanded from a local tool to a multi-channel platform (iMessage, Raft agent network, Telegram, desktop) with five sandboxed execution backends P1W6. Model releases span the parameter spectrum: 405B down to 1B, with specialized variants for tool-calling, financial prediction, and code (NousCoder-14B) E7E15E39E40E43. The Atropos RL framework shipped alongside tinker-atropos for managed training, indicating a buildout pipeline where environments → RL training → specialist model → deployment is being systematized P21P26E33.

Research themes

1. Long-context efficiency — Lighthouse Attention demonstrates a selection-based hierarchical attention for pretraining, targeting 1.4–1.7× speedup at scale without custom sparse kernels or auxiliary losses W3W4. The scaling-transformer repo (Jul 2025) and infini-attention repo (Jul 2024) suggest sustained attention-architecture exploration P8P20. 2. Reinforcement learning for LLMs — Atropos is a dedicated LLM RL gym, and the DeepHermes Atropos specialist models (ToolCalling, Financial, AscensionMaze-RLAIF) show an end-to-end RL training pipeline producing task-specialized models P21E39E43E54. The NeMo RL and NeMo Gym forks (Apr 2026) indicate deeper RL infrastructure investment E59E60. 3. Distributed training — DisTrO/Psyche Network claims 3–4 orders of magnitude reduction in inter-GPU communication, demonstrated by training a Consilience 40B model P10. Production code lives in the PsycheFoundation org in Rust (distro.rs) P10. 4. Reasoning and math — Open-Reasoning-Tasks (493 stars) curates reasoning task collections; Nomos (194 stars) is a reasoning harness for mathematical proof-writing with self-critique loops; Forge API offered reasoning completions at fast/medium/slow presets P6P14P25. 5. Agent self-improvement — hermes-agent-self-evolution (4,364 stars) uses DSPy + GEPA to optimize agent skills, prompts, and code autonomously E36. 6. MoE architectures — The moe-10b-a1b-8k-wsd-lr3e4-1t model (10B total, ~1B active) shows experimentation with mixture-of-experts at modest scale E45. 7. Alternative-language tooling — Kaida (Kotlin LLM library), wandb-rs (Rust Weights & Biases client), solana-flake (Nix/Solana build tooling), and iroh-fake-store (Rust blob store testing) indicate polyglot infrastructure investment P11P12P17P27.

Hiring & scaling

Six open roles as of June 2026 paint a picture of productization + enterprise maturity: UI/UX Designer and Full Stack Engineer for consumer-grade Hermes Desktop surfaces; Forward Deployed Engineer for enterprise customer onboarding; General Counsel signaling legal/commercial readiness for platform-scale operations; Research Scientist and Machine Learning Engineer sustaining model R&D E19E20E21E22E23E24. No cited evidence on team sizes, locations, or compensation bands — the job listings lack descriptions beyond titles and the ATS source marker.

The fork pattern strongly implies a NVIDIA partnership or heavy NeMo/Megatron dependency: 9 of the 12 cited forks (May 2026 cluster) are NVIDIA repos covering the full training stack — Megatron-LM (distributed training), Megatron-Bridge (NeMo adapter), Automodel (model selection), NeMo RL/Gym (RL infrastructure), Nemotron (model architecture), OpenShell/OpenShell-Community (terminal), NemoClaw (agent shell tools) E28E29E30E37E46E47E50E51E52E58E59E60. Combined with NVIDIA/skills becoming a built-in trusted skills tap in Hermes Agent W2, this suggests a deepening technical integration.

Category implications

Infrastructure — The Megatron-LM + Megatron-Bridge + NeMo RL forks indicate Nous is building or inspecting large-scale training infrastructure on NVIDIA's stack. Combined with DisTrO's distributed training claims and the Psyche Network, Nous is investing in training infrastructure that could reduce dependency on centralized compute providers E29E30E59E60P10. The wandb-rs Rust client and iroh-fake-store suggest in-house MLOps tooling in Rust P12P27.

Product — Hermes Agent is the primary product vector: a multi-channel (desktop, iMessage, Telegram, Raft), multi-model (300+ via Portal), self-improving agent platform with 190k+ GitHub stars P1P24W6. The paperclip-adapter ("run Hermes as a managed employee in a Paperclip company") and agent-governance-toolkit fork point toward enterprise agent deployment and governance as a GTM motion E37E55.

Research — RL is the dominant research theme, with Atropos as gym, specialist models as outputs, and thinking-machines Tinker integration for managed training P21P26E39E43E54. Long-context attention (Lighthouse) and MoE architectures represent efficiency research bets W3W4E45. Open-Reasoning-Tasks and Nomos anchor the reasoning research track P6P25.

Hiring — The mix of product roles (UI/UX, Full Stack, Forward Deployed) alongside research roles (Research Scientist, MLE) and a General Counsel indicates a transition from pure research lab to product company with enterprise sales motion E19E20E21E22E23E24.

GTM — The Nous Portal aggregates 300+ models at discounted rates, positioning Hermes-4 models as a funnel into the platform W1. Hermes Desktop's free MIT-licensed public preview serves as a top-of-funnel acquisition channel. The Forward Deployed Engineer role signals enterprise deployment deals E22. The General Counsel hire indicates contracts, compliance, and possibly revenue-generation maturity E20. No cited evidence of revenue figures, pricing tiers beyond "heavily discounted rates," or paying customer counts.

Traction highlights

• hermes-agent: 190,109 stars, 32,944 forks, 19,913 open issues — top-tier open-source traction rivaling major frameworks P24. The HN event at repo creation scored only 4 points/3 comments E1, but subsequent growth is extraordinary.
• hermes-agent-self-evolution: 4,364 stars since March 2026 E36.
• hermes-paperclip-adapter: 1,650 stars since March 2026 E55.
• atropos: 1,273 stars, 367 forks since April 2025 P21E33.
• DisTrO: 1,034 stars, 88 points/26 comments on HN at launch P10E3.
• Hermes-Function-Calling: 1,386 stars, 198 forks P9.
• Open-Reasoning-Tasks: 493 stars, 72 forks P6.
• Hermes-3-Llama-3.1-8B: 312,589 HuggingFace downloads, 464 likes E2.
• Hermes-3-Llama-3.1-70B: 131,870 downloads E13.
• Hermes-4-14B: 31,365 downloads E11.
• Hermes-4.3-36B: 25,308 downloads E6.

Community contribution volume is notable: v0.17.0 alone had 245 contributors and 300+ issues closed P1.