Nous Research

Thesis

Nous Research is moving from a post-training-focused open-weights lab into a vertically integrated, open-source agent platform. The pack shows a continuing model-release line (Hermes 4 family, NousCoder, nomos-1) alongside a far denser stream of Hermes Agent engineering: a near-weekly stable release cadence E8E10E12E20E24E34E37E40E41E43, a plugin ecosystem consolidating into one first-party repo P3, and eight memory providers spun out as standalone handoffs P18P19P20P21P22P23P24P25. The funding posture — a $90M Series B at a $1.5B valuation earmarked for 'Hermes for Businesses' W6 — plus the company's own framing that 'tomorrow's moats will be built around execution, trust, and extensibility' W1 indicate the strategic bet has shifted from weights toward agent distribution, plugin extensibility, and enterprise deployment.

Signal desks

Hiring:

• Six roles opened June 5, 2026 on the careers page: UI/UX Designer, General Counsel, Full Stack Engineer, Forward Deployed Engineer, Research Scientist, and Machine Learning Engineer E55E56E57E58E59E60.
• A third-party account at the same period counted 13 open roles spanning security, infrastructure, evaluation, product engineering, mobile development, forward-deployed engineering, analytics, support, product management, finance, and office operations W5.
• Forward Deployed Engineer plus the 'Hermes for Businesses' push point to a commercialization buildout E58W6.
• Thin on specifics: no job descriptions or locations are present in this pack, so team/hub granularity beyond role titles is not evidenced.

Forks:

• A cluster of NVIDIA-NeMo and NVIDIA agent/training repos are forked: Gym, Automodel, RL, Nemotron, NemoClaw, and OpenShell E47E48E49E50E51E52, signaling inspection of RL and agent-training/agent-shell infrastructure.
• Nous forked hexgrad/misaki, the G2P engine for Kokoro TTS models (Python, Apache-2.0) P12E19.
• Other forks: jim-plus/llm-abliteration (156 stars) E45, microsoft/agent-governance-toolkit E46, harbor-framework/harbor E53, cline/cline E54, and getitcheappro/image-size-patched E44.
• Language is only evidenced for misaki (Python) P12; fork events otherwise carry repo/parent metadata only.

Releases:

• Hermes Agent shipped a steady sequence of stable tags from v2026.8.27 through v0.22.0 (the 'Threshold Release') E37E40E41E43E34E24E20E12E10E8.
• v0.22.0 is the headline: guided first-launch setup on a free trial model, app connectors with no account needed, a one-click llama.cpp local runtime, signed desktop packages, a 600-plugin catalog, and a codebase a third smaller P1.
• Hugging Face model releases: Hermes-4-405B, Hermes-4-70B, Hermes-4-14B, Hermes-4.3-36B, NousCoder-14B, and nomos-1 E2E3E4E5E6E7.
• Memory-provider plugins were cut as standalone releases: retaindb v1.0.1, byterover v1.0.1/v1.0.2, and holographic v0.1.1 P8P9P10P11.

Talking:

• The fundraise note frames Nous's lane as post-training ('the layer where we could have the biggest impact') and Hermes as a learning agent released under MIT, with moats in execution, trust, and extensibility W1.
• The Hermes Index announcement introduces an average of four benchmarks — Hermes Bench, Terminal-Bench 4.0, Terminal-Bench-Science, and SkillsBench — to inform model choice and 'show labs how they perform in Hermes' W4.
• MISSINGNO is announced as a free, token-efficient coding and agentic reasoning model on Nous Portal (listed as 'stealth/missingno' at $0.00/M tokens); open-weight status is unconfirmed W3.
• Press centers the $90M round and the enterprise launch ('Hermes for Businesses'), with Samsung (and Nvidia per SiliconANGLE) among backers W2W6.

Shipping

The dominant shipping artifact is Hermes Agent. v0.22.0 (Oct 13, 2026) is the largest window — described as ~2.5x the v0.21.0 cycle — adding a self-setting-up first launch, connectors, one-click llama.cpp local models, tested SemVer stable releases, signed desktop packages including the Microsoft Store, a 600-plugin catalog, and the Claude 5.5 / GPT-6 model wave P1. The immediately prior patches show platform maturation: stable-channel update tracking P2, a new stable release pipeline producing tested Docker images P5, and a broad plugin-SDK/Desktop/connectors wave in v0.21.5 P13. Patch velocity is high: v0.21.4 rolled up ~1,800 PRs and v0.21.5 ~460 PRs, both flagged as stable tags for downstream consumers P16P13.

On the ecosystem side, first-party plugins are consolidating into NousResearch/hermes-official-plugins ('one repo, every first-party plugin') P3, with individual plugin repos (Spotify, Home Assistant, Snyk, Blender, NVIDIA, TouchDesigner, accent-picker, Claude subscription) marking 'Moved to hermes-official-plugins' E11E14E35E21E36E38E23. Separately, eight memory providers were extracted to standalone 'handoff' repos so upstream maintainers can take them over P18P19P20P21P22P23P24P25.

On the model side, Hugging Face shows a line of first-party finetunes: the Hermes 4 family (405B, 70B, 14B, 4.3-36B), NousCoder-14B, and nomos-1 — Apache-2.0 except the Llama-3-licensed 70B/405B E2E3E4E5E6E7.

Research themes

• Post-training and synthetic data. Nous frames post-training as its highest-impact layer W1. The Hermes 4 series was customized with a dataset of 5M synthetically generated records, about two-thirds aimed at reasoning, generated with a custom tool called DataForge W2.
• Agentic memory. The densest research signal is memory: holographic (SQLite FTS5 + trust scoring + HRR), hindsight (knowledge graph + entity resolution), honcho (dialectic Q&A user modeling), mem0, openviking, retaindb, supermemory, and byterover P18P19P20P21P22P23P24P25, alongside a read-only Memory Wiki dashboard auditing state.db and MEMORY.md/USER.md entries P26.
• RL and agent training infrastructure. A run of NVIDIA-NeMo forks (RL, Gym, Automodel, Nemotron) plus NemoClaw and OpenShell indicates active inspection of RL and agent/training tooling E47E48E49E50E51E52.
• Self-improvement. A dedicated repo, hermes-agent-self-evolution (5,502 stars), describes evolutionary self-improvement via DSPy + GEPA E39, consistent with Hermes's 'writes skills from experience' framing P3.
• Evaluation. Hermes Index averages four benchmarks, including the in-house Hermes Bench W4.
• Speech/TTS. The misaki fork is a G2P engine designed for Kokoro models P12.
• Coding models. NousCoder-14B E3 and the MISSINGNO coding/agentic-reasoning model W3 extend the coding-model line.

Hiring & scaling

The earliest concrete hiring evidence is a six-role opening wave on June 5, 2026: UI/UX Designer, General Counsel, Full Stack Engineer, Forward Deployed Engineer, Research Scientist, and Machine Learning Engineer E55E56E57E58E59E60. A contemporaneous account tallies 13 open roles across security, infrastructure, evaluation, product engineering, mobile, forward-deployed engineering, analytics, support, product management, finance, and office operations W5. The mix implies parallel scaling of research (Research Scientist, ML Engineer) and commercialization (Forward Deployed Engineer, General Counsel, product/UI roles) E55E56E57E58E59E60. The Series B is explicitly directed at the enterprise push, so hiring is likely to keep tilting toward deployment and product W6. No location data or role-level descriptions are evidenced in this pack — that gap itself is the notable limitation.

Category implications

• GTM. The clearest category shift is to enterprise: the $90M Series B funds 'Hermes for Businesses,' positioned for companies deploying customized agents while keeping data private W6. This aligns with a Forward Deployed Engineer role E58 and the framing of an open-source enterprise agent whose wager is that user control is an advantage rather than a tax W5.
• Product / platform. Consolidation into hermes-official-plugins P3 and the 600-plugin catalog P1 show a deliberate move to a plugin/ecosystem model — shipping integrations (Home Assistant, Spotify, Blender, NVIDIA, Snyk, Claude subscription) as separable first-party packages rather than bundling them in core P4P7P14P15P17P28.
• Infrastructure. Hermes is pushing toward a distributable, self-hostable platform: stable SemVer channels, signed packages, Docker/Hermes Cloud images, and a one-click llama.cpp local runtime P1P2P5. The tirith prototype shows command-security scanning tooling being considered, then archived as unmaintained P6.
• Research. The synthetic-data/DataForge pipeline W2 and the NVIDIA-NeMo RL/training forks E48E49E50 imply continued investment in post-training and RL-based agent capability, not only wrapper engineering.
• Evals. Hermes Index (Hermes Bench + Terminal-Bench 4.0 + Terminal-Bench-Science + SkillsBench) signals an effort to own the evaluation surface for agentic models — positioning Nous to 'show labs how they perform in Hermes' W4.
• Memory/data. Spinning eight memory backends out as standalone handoffs P18P19P20P21P22P23P24P25 indicates a deliberate de-bundling of memory into an extensible provider layer — consistent with the extensibility-moat thesis W1 — rather than betting on a single proprietary memory stack.

Traction highlights

• Hermes has been downloaded more than 24M times, with several thousand open-source contributors and an estimated 2.5% of world token consumption W2.
• The hermes-agent repo shows 253,584 stars E1.
• Nous raised a $90M Series B at a $1.5B valuation, led by Robot Ventures with Samsung participation W6, with Nvidia and Samsung named as backers in coverage W2.
• Hermes Agent v0.22.0 rolls up ~24,200 commits, ~6,080 merged PRs, ~2,520 contributors, and ~7,180 closed issues in one window P1.
• Hugging Face traction is modest but real: Hermes-4-14B leads downloads (2,808), Hermes-4.3-36B leads likes (307) E2E5.
• Individual plugin repos show early pull: claude-subscription-directsdk 122 stars E23, hermes-official-plugins 271 stars E9, backsearch 64 stars E42, and hermes-memory-wiki 34 stars E33.