Nous Research analysis

Thesis

Nous Research is in the middle of a visible pivot from an open-weights post-training lab into an agent-platform company. The fundraise narrative is explicit on both halves: post-training was "the layer where we could have the biggest impact" via fine-tuning Meta's Llama models, and Hermes Agent — released under MIT in February — is framed as a self-improving agent whose "moats will be built around execution, trust, and extensibility" W1. The Series B is being spent on exactly that: "Hermes for Businesses," an enterprise agent push W2 W5. The repository evidence confirms the platformization thesis — first-party capabilities are being extracted out of Hermes core into a catalog of standalone plugins (Spotify, Home Assistant, NVIDIA, Blender, Claude Subscription DirectSDK, Snyk, TouchDesigner) P1 P2 P5 P12 P13 P15 P26 P27, memory providers are being spun out as "handoff" repos for external maintainers P16 P17 P18 P19 P20 P21 P22 P23, and a newly created "stable release pipeline" now cuts tagged releases for Docker and Hermes Cloud P3. The lab is building distribution, extensibility, and enterprise readiness rather than running a purely model-centric cadence.

Signal desks

Hiring — Captured open roles: UI/UX Designer, General Counsel, Full Stack Engineer, Forward Deployed Engineer, Research Scientist, Machine Learning Engineer E55 E56 E57 E58 E59 E60. A contemporaneous snapshot lists 13 open roles across security, infrastructure, evaluation, product engineering, mobile development, forward-deployed engineering, analytics, support, product management, finance, and office operations W6. The mix of General Counsel, Forward Deployed Engineer, support, finance, and office operations reads as an enterprise-GTM buildout, and mobile development tracks the stated mobile-app goal W4 W6. Caveat: the pack carries role titles only — no locations, team ownership, or job-description detail is cited, so the hiring signal is thematic rather than granular E55 E56 E57 E58 E59 E60.
Forks — A heavy NVIDIA/NeMo cluster: NemoClaw, Gym, Automodel, RL, Nemotron, and OpenShell E47 E48 E49 E50 E51 E52, indicating active inspection of NVIDIA's agentic RL and post-training tooling. Also forked: microsoft/agent-governance-toolkit E46, cline/cline E54, harbor-framework/harbor E53, llm-abliteration E45, image-size (patched) E44, and hexgrad/misaki — the Kokoro G2P/TTS engine with Japanese pitch-accent support P10 E17. The repeated NVIDIA forks are the clearest forward-looking signal (training/agent RL infra); misaki points to TTS/audio interest P10.
Releases — Models: the Hermes-4 series (405B, 70B, 14B, 36B), NousCoder-14B, and nomos-1 E2 E3 E4 E5 E6 E7. Software: a fast Hermes Agent patch cadence (v0.21.2 through v0.21.6 across Sept–Oct 2026) P3 P11 P14 P25 P28, plus first catalog releases for the extracted memory providers byterover, retaindb, and holographic P6 P7 P8 P9. MISSINGNO, a coding/agentic-reasoning model, is announced as launching on Nous Portal and Hermes Agent W4.
Talking — The fundraise note frames post-training as the point of leverage and Hermes Agent as "the more you use it, the better it gets" W1. Enterprise/control framing dominates press: "Hermes for Businesses" with private, secure multi-step workflows W2; open-source architecture as a way to gain data control and reduce vendor dependence W5 W6. Quantified traction claims: 24M+ downloads, ~2.5% of the world's token consumption, thousands of contributors, and the DataForge synthetic dataset (5M records, ~two-thirds reasoning-targeted) W3. MISSINGNO plus a mobile app and the investor roster (Nvidia, M12, Samsung Next, Robot Ventures, USV, YC, Menlo) are also in the public narrative W4.

Shipping

What actually shipped in this pack is overwhelmingly Hermes platform infrastructure, not new frontier models. The v0.21.6 tag is the first cut by the new stable release pipeline — one attempt ref, a tested Docker image, and a receipt tag — rolling up ~2,106 merged PRs and ~8,867 non-merge commits since v0.21.5 P3. Earlier patch tags document the same machinery: ~460 PRs in v0.21.5 P11, ~1,812 PRs in v0.21.4 P14, ~338 PRs in v0.21.3 P25, and ~312 PRs in v0.21.2 P28. First-party plugins consolidated into a single hermes-official-plugins repo: Home Assistant, Spotify, Claude Subscription DirectSDK "and more" P1 E8, with individual plugin repos for Spotify P2, Home Assistant P5, NVIDIA App/Broadcast P12, Blender P13, the Claude Pro/Max DirectSDK provider P15, Snyk P26, and TouchDesigner P27. Ten-plus memory providers were cut from core into standalone handoff/plugin repos (byterover, holographic, hindsight, honcho, openviking, mem0, retaindb, supermemory) P16 P17 P18 P19 P20 P21 P22 P23, with a read-only Memory Wiki dashboard for auditing state.db and persistent memory P24. On models, the Hermes-4 family, NousCoder-14B, and nomos-1 predate this window but remain the released first-party fine-tunes in evidence E2 E3 E4 E5 E6 E7. Notably, hermes-plugin-tirith shipped as an archived, explicitly unmaintained prototype P4 E11.

Research themes

Post-training / fine-tuning as the core competency — Hermes-4 405B/70B/14B/36B, NousCoder-14B, and nomos-1 are all first-party fine-tunes E2 E3 E4 E5 E6 E7, consistent with the stated "post-training was the layer where we could have the biggest impact" framing W1.
Synthetic data at scale — Hermes 4 was customized with ~5M synthetically generated records, ~two-thirds aimed at reasoning, via a custom tool called DataForge W3.
Agentic RL / training infrastructure — Repeated forks of NVIDIA NeMo RL/Gym/Automodel/Nemotron and NemoClaw/OpenShell signal active work on agent training and RL tooling E47 E48 E49 E50 E51 E52.
Self-improving agents — Hermes Agent is described as "the agent that grows with you" E1, and hermes-agent-self-evolution ("optimize skills, prompts, and code using DSPy + GEPA") is the highest-starred research-adjacent repo in the pack (5,490 stars) E38.
Memory architectures — The memory-provider handoffs expose a broad survey of approaches: SQLite + FTS5 + HRR compositional retrieval (holographic) P17, knowledge-graph + entity resolution + multi-strategy retrieval (hindsight) P18, cross-session user modeling with dialectic Q&A (honcho) P19, filesystem-style context DBs (openviking) P20, and cloud hybrid vector+BM25+rerank (retaindb) P22.
TTS/G2P — The misaki fork (Kokoro G2P, Japanese pitch-accent) is the only audio/G2P signal in the pack P10 E17.

Hiring & scaling

Hiring evidence is thin on detail but directionally clear. Six roles were captured from the careers page — UI/UX Designer, General Counsel, Full Stack Engineer, Forward Deployed Engineer, Research Scientist, Machine Learning Engineer E55 E56 E57 E58 E59 E60 — and a publication snapshot counted 13 open roles including security, infrastructure, evaluation, product engineering, mobile development, forward-deployed engineering, analytics, support, product management, finance, and office operations W6. This maps to a Series-B scaling wave: enterprise GTM (FDE, General Counsel, support, finance), a mobile app buildout W4, and continued research/infra/eval hiring W6. No locations, team structures, or job descriptions are cited, so headcount, hub strategy, and infrastructure spend cannot be inferred beyond role titles E55 E56 E57 E58 E59 E60 W6.

Category implications

Strategy — The open-license, user-control bet is the stated thesis: moats built on "execution, trust, and extensibility" W1, with open-source positioned as more data control and less vendor lock-in for enterprises W5 W6.
Infrastructure — Extracting memory providers into standalone plugins P16 P17 P18 P19 P20 P21 P22 P23 and adding per-profile plugin isolation (plugins.isolation: host) P3 signals a durable plugin/extension architecture replacing the old monolithic core. The stable release pipeline for Docker/Hermes Cloud is explicit release-engineering investment P3.
Product — First-party plugins span consumer surfaces (Spotify, Home Assistant) and professional/creative tooling (Blender, TouchDesigner, NVIDIA App/Broadcast, Snyk) P1 P2 P5 P12 P13 P26 P27; the Claude Subscription DirectSDK plugin routes the Claude Pro/Max subscription through the official Claude Code CLI as a model provider P15. The stated product frontier is "Hermes for Businesses" W2 plus a mobile app W4.
Research — Post-training and synthetic data (DataForge) remain the backbone W1 W3; the NVIDIA NeMo RL forks and the DSPy+GEPA self-evolution repo point toward agentic training and self-improvement as the research direction E38 E48 E49 E50 E51.
Hiring — Role mix implies an enterprise-GTM and product buildout (General Counsel, Forward Deployed Engineer, mobile, support, finance, office ops) layered on continued security/infra/eval/research staffing E56 E58 W4 W6.
GTM — The Series B is explicitly directed at the enterprise sector W2 W5. No revenue, customer-count, or pricing figures are cited in this pack, so specific commercial claims are unsupported W2 W3 W4 W5 W6.

Traction highlights

$90M Series B at a $1.5B valuation W2 W5.
24M+ Hermes downloads and an estimated ~2.5% of the world's token consumption W3.
NousResearch/hermes-agent at ~252,609 GitHub stars E1.
140 contributors appearing in the single v0.21.2 release window P28.
Model pull on Hugging Face: e.g., Hermes-4-14B at 2,099 downloads / 191 likes E5, Hermes-4.3-36B at 1,646 downloads / 306 likes E2, Hermes-4-70B at 1,412 downloads / 223 likes E4.
hermes-agent-self-evolution at 5,490 stars signals strong community interest in the self-improvement direction E38.