Nous Research

Thesis

Nous Research is executing a decisive pivot from open-weight model publisher to integrated agent platform company. Throughout mid-2026, the overwhelming majority of engineering velocity has concentrated on hermes-agent — a conversational AI agent that ships near-weekly releases, draws 650+ contributors, and has accumulated over 227k GitHub stars E1P3. Model releases continue (NousCoder-14B, Hermes-4.3-36B, nomos-1) but increasingly serve the agent ecosystem: tool-calling specialists, RL-trained variants, and coding models that complement the agent harness E6E4E11E34. The hiring of a General Counsel, Forward Deployed Engineers, and a UI/UX Designer — alongside continued Research Scientist and MLE openings — signals an active commercialization buildout E39E40E41E42. The fork pattern reveals deep investment in the NVIDIA NeMo ecosystem for RL training infrastructure and eval rigor, while the Telegram Business plugin, Paperclip adapter, and subscription management features point toward emerging SMB and enterprise GTM motions P10E48P7.

Signal desks

Hiring

• Research Scientist — continuing investment in core model R&D and post-training capabilities E43
• Machine Learning Engineer — generalist MLE role, plus a separate externally advertised MLE, Evals role focused on agent capability evaluations, benchmark design, and infrastructure development supporting researchers E44W3
• Full Stack Engineer — platform and product engineering to support the hermes-agent surface area (CLI, desktop, gateway) E41
• Forward Deployed Engineer — classic commercialization signal; implies customer-facing integration work, likely tied to the Telegram Business and Paperclip enterprise adapter offerings E42
• UI/UX Designer — dedicated design hire for agent interfaces; aligns with the desktop app platformization, voice mode UX, and consumer-facing product polish described in recent releases E39
• General Counsel — strong signal of legal/regulatory maturity, revenue contracts, and potential fundraising or partnership negotiations underway E40

Forks

• vllm-project/speculators (Python) — speculative decoding draft model training library; signals investment in inference speed optimization for deploying Hermes models at scale E32P13
• harbor-framework/harbor — agent evaluation framework from the Terminal-Bench creators; used for benchmarking agents like Claude Code and OpenHands at scale; implies Nous is evaluating hermes-agent against competitive agent frameworks E29P11
• NVIDIA-NeMo/Automodel, NVIDIA/Megatron-LM, NVIDIA-NeMo/Megatron-Bridge (Python) — simultaneous fork of three NVIDIA training infrastructure repos on 2026-05-27; strongly signals large-scale training or fine-tuning pipeline buildout using NeMo as the training backbone E51E52E53
• NVIDIA/NemoClaw, NVIDIA-NeMo/Gym, NVIDIA-NeMo/RL — additional NeMo ecosystem forks for RL training environments and agent RL workflows; complements the Atropos RL framework E58E59E60
• pytorch/torchtitan — PyTorch's native large-scale training framework; reinforces the distributed training infrastructure theme alongside the NeMo forks E57
• microsoft/agent-governance-toolkit — agent safety and governance evaluation tooling from Microsoft; suggests Nous is building internal agent safety/alignment evaluation capabilities E56
• jim-plus/llm-abliteration — LLM refusal/alignment ablation techniques; consistent with Nous's "unrestricted" model philosophy E55
• getitcheappro/image-size-patched (TypeScript) — security-patched image-size fork for denial-of-service fixes in image parsers; likely a supply-chain security dependency for hermes-agent or related tooling E15P1

Releases

• hermes-agent v0.20.0 ("The Herald Release," 2026-08-03) — streaming conversational voice with barge-in, Agent-to-Agent (A2A) v1.0 protocol support, signed outbound webhooks, grounded research with verifiable citations, plugin SDK, and artifact previews. ~3,650 commits, ~1,400 merged PRs, 650+ contributors P3E17W2W4
• hermes-agent v0.19.1 (2026-07-30) — patch release rolling up ~1,000 PRs into a stable tag for Docker images and hosted deployments; dominated by bug-fix and salvage waves across gateway, voice subsystem, desktop, and installer P4E18
• hermes-agent v0.19.0 ("The Quicksilver Release," 2026-07-20) — ~80% reduction in first-turn time-to-first-token, reasoning streams live by default, desktop app 20-PR speed overhaul, Nous subscription management from terminal, Bitwarden/1Password integration, smart approvals for flagged commands, durable delivery ledger. ~2,245 commits, ~1,065 merged PRs, ~3,300 issues closed, 450+ contributors P7E21
• hermes-agent v0.18.2 and v0.18.1 (2026-07-07/08) — patch releases fixing WhatsApp Baileys dependency and rolling up ~660–667 PRs of bug fixes and stability work P8P9E24E25
• hermes-agent v0.18.0 ("The Judgment Release," 2026-07-01) — complete P0/P1 backlog sweep: 692 highest-priority items resolved, zero open critical/high issues. Mixture-of-Agents as first-class citizen, self-verification against evidence, completion contracts for /goal, scale-to-zero gateway, fan-out background subagents. ~1,720 commits, 998 merged PRs, 949 issues closed, 370+ contributors P12E30
• hermes-agent v0.17.0 ("The Reach Release," 2026-06-19) — iMessage via Photon Spectrum, Raft agent network integration, Cursor Composer model via xAI Grok subscription, subagent background execution, dashboard profile builder, Skills Hub, memory tool upgrade. ~1,475 commits, ~800 merged PRs, 245 contributors P14E35
• Model releases: NousCoder-14B (67.87% LiveCodeBench v6, Apache 2.0) E6W1; Hermes-4.3-36B (14,890 downloads, Apache 2.0) E4; nomos-1 (30.5B params, Apache 2.0) E11; Hermes-4-405B E14; Hermes-4-70B E7; Hermes-4-14B E9; Minos-v1 (text-classification) E10; moe-10b-a1b-8k-wsd-lr3e-1t (10.5B MoE) E36; k2-merged-3.5T-bf16 (3.47T params) E45; multiple DeepHermes-3 and Atropos-trained specialist variants E3E13E19E34E37E46
• Infrastructure releases: hermes-toolperf-evals (core-toolset A/B eval harness from 1.5M-message session DB mining with NeMo Relay traces) P2E16; hermes-agent-ci-infra (GKE self-hosted GitHub Actions runners, idle ~$25/mo) P6E22; hermes-e2e-evidence (CI visual evidence publishing) P5E20

Talking

• Agent platform as primary narrative — Nous frames hermes-agent as the center of gravity: "Hermes is the herald of the gods" with voice, A2A protocol, webhooks, and citations as marquee features. The release notes are detailed, product-focused, and emphasize contributor scale (650+ contributors, thousands of issues closed) P3W2W4
• Open coding model positioning — NousCoder-14B is framed as an open alternative for competitive programming in the "Claude Code moment," with the Atropos RL framework, training harness, and benchmarks released alongside weights W1E6
• "Unrestricted" and "human-centric" brand — LinkedIn and third-party coverage consistently describe Nous as building "personalized, unrestricted" AI and "open-source, human-centric" models, distinguishing from OpenAI and Anthropic's content policies W6
• Workforce scale — publicly described as 11–50 employees, headquartered in New York W3; co-founder "Teknium" (Ryan) identified as head of post-training W4
• Limited direct long-form publishing in this pack — most talking evidence comes from release notes, LinkedIn posts, and third-party coverage rather than research blogs or papers; the analysis depends primarily on GitHub artifacts and release changelogs for signal W2W4W1

Shipping

Nous's shipping velocity is defined by the hermes-agent release cadence. From late May through early August 2026, the team shipped at least 11 tagged releases — roughly one every 6–7 days P3P4P7P8P9P12P14E17E18E21E24E25E30E35E38E47E49E50. v0.20.0 alone encompassed ~3,650 commits and ~1,400 merged PRs from 650+ contributors P3. The model release side continues but at lower tempo: NousCoder-14B, Hermes-4.3-36B, nomos-1, and multiple Atropos-trained specialist models shipped across the same window E6E4E11E34E37E46. Infrastructure tooling shipped in parallel: hermes-toolperf-evals for agent tool-calling evaluation, hermes-agent-ci-infra for GKE-based self-hosted CI, and hermes-telegram-business for SMB automation P2P6P10.

Research themes

1. Agent RL and post-training — The Atropos framework (1,344 stars) provides RL environments for collecting and evaluating LLM trajectories, with multiple DeepHermes specialist models trained through it (ToolCalling, Financial-Fundamentals, AscensionMaze-RLAIF) E28E34E37E46. The NVIDIA NeMo ecosystem forks (Automodel, Megatron-LM, Megatron-Bridge, NemoClaw, Gym, RL) suggest large-scale RL training infrastructure buildout E51E52E53E58E59E60. 2. Tool-calling evaluation and optimization — The hermes-toolperf-evals repo contains a 9-case A/B eval harness built from production session-DB mining (~1.5M messages) and NeMo Relay traces, targeting specific tool-use failure modes (ambiguous edits, hidden searches, big output handling, etc.) P2. This connects to a 15-PR tool-efficiency improvement tracker in hermes-agent P2. 3. Speculative decoding for inference — The speculators fork (vllm-project) signals active work on draft-model speculative decoding to reduce LLM inference latency, relevant for serving Hermes models at scale E32P13. 4. Distributed training — DisTrO (1,034 stars) demonstrated 3–4 order-of-magnitude reduction in inter-GPU communication, culminating in the Psyche Network and a 40B Consilience model P23E2. The torchtitan fork reinforces continued attention to large-scale distributed training E57. 5. Agent self-improvement — hermes-agent-self-evolution (4,960 stars) uses DSPy + GEPA for evolutionary optimization of skills, prompts, and code, indicating a meta-learning approach to agent capability improvement E31. 6. Agent governance and safety — The microsoft/agent-governance-toolkit fork suggests internal work on agent alignment evaluation, while the llm-abliteration fork aligns with Nous's "unrestricted" model philosophy E56E55.

Hiring & scaling

Six roles advertised via Nous's careers page as of early June 2026, plus one external MLE Evals listing E39E40E41E42E43E44W3:

| Role | Signal | |---|---| | Research Scientist | Continued core research investment E43 | | Machine Learning Engineer | Generalist MLE; separate Evals-focused MLE for agent benchmark design and eval infra E44W3 | | Full Stack Engineer | Platform/product buildout around hermes-agent E41 | | Forward Deployed Engineer | Customer integration, enterprise deployment — commercialization signal E42 | | UI/UX Designer | Consumer/business product polish for voice UX, desktop app, dashboard E39 | | General Counsel | Legal maturity, contracts, fundraising, regulatory readiness E40 |

The mix of Research Scientist + MLE alongside Forward Deployed Engineer + General Counsel + UI/UX Designer indicates an organization simultaneously investing in core research and preparing for revenue-scale operations. Publicly described headcount is 11–50 employees, New York-based W3. The hermes-agent contributor count (650+ on v0.20.0) vastly exceeds internal headcount, reflecting an open-source contribution model P3.

Category implications

Strategy: Nous is repositioning from an open-weight model publisher (the Hermes fine-tune series) to an agent platform company. The hermes-agent release frequency dwarfs model release frequency, and every model release now ties back to agent capabilities — coding (NousCoder), tool-calling (Atropos specialists), and reasoning (DeepHermes) P3P7W1E6E34. The Paperclip adapter ("run Hermes as a managed employee") and Telegram Business plugin signal a two-pronged GTM: consumer agent + SMB/enterprise automation E48P10.

Infrastructure: The cluster of NVIDIA NeMo ecosystem forks (Megatron-LM, Automodel, Megatron-Bridge, NemoClaw, Gym, RL) plus torchtitan indicates substantial training infrastructure investment, likely for RL-based post-training at scale E51E52E53E58E59E60E57. The GKE-based self-hosted CI infrastructure (hermes-agent-ci-infra) with spot instances and GCS-backed caching suggests operating at a scale where managed GitHub Actions runners are cost-prohibitive P6.

Product: Hermes Agent has evolved into a multi-surface platform: CLI, desktop app (with plugin SDK and artifact previews), gateway, and messaging channels (iMessage via Photon, Telegram, WhatsApp) P3P14P7P8P10. v0.20.0 added streaming voice with barge-in, A2A protocol support, and signed outbound webhooks — positioning Hermes as an interoperable, persistent agent rather than a chat interface P3W4. Subscription management from the terminal confirms a monetization path P7.

Research: Research is tightly coupled to product. The Atropos RL framework directly produces specialist models deployed in hermes-agent E28E34. The hermes-toolperf-evals repo mines production data to drive tool-calling improvements P2. The harbor-fork suggests competitive benchmarking against other agent frameworks (Claude Code, OpenHands) E29. This is applied research with near-immediate product impact.

Hiring: The Forward Deployed Engineer role, combined with Telegram Business and Paperclip plugins, indicates an enterprise deployment motion forming E42P10E48. The General Counsel hire suggests Nous anticipates contracts, partnerships, or regulatory engagement that requires dedicated legal counsel E40. The MLE Evals role points to systematic agent evaluation as an organizational priority W3.

GTM: Distribution flows through open-source (227k+ GitHub stars on hermes-agent E1), messaging platforms (iMessage, Telegram, WhatsApp, Raft) P14P10, and a subscription model managed in-terminal P7. The strategy appears to be: maximize reach through every chat surface, convert power users to subscriptions, and serve businesses through managed agent deployments (Paperclip, Telegram Business) E48P10.

Traction highlights

• hermes-agent: 227,626 GitHub stars, 650+ contributors on v0.20.0 alone E1P3
• hermes-agent-self-evolution: 4,960 stars E31
• Hermes-Function-Calling: 1,386 stars P22
• Atropos (RL environments framework): 1,344 stars E28
• DisTrO: 1,034 stars, HN traction (88 points, 26 comments) P23E2
• Model downloads: Hermes-3-Llama-3.2-3B at 6,018 downloads E8; Hermes-4.3-36B at 14,890 downloads E4; Hermes-4-14B at 4,845 downloads E9
• Release velocity: 11+ tagged hermes-agent releases in ~70 days (late May to early August 2026) P3P4P7P8P9P12P14
• Issue resolution throughput: ~700 P0/P1 issues cleared in 12 days for v0.18.0; ~3,300 issues closed in v0.19.0 P12P7