Nous Research shipped Hermes Agent v0.20.0 on August 3, 2026 - the biggest single release in the project's history by its own changelog numbers. The four headline items are real-time voice, an Agent-to-Agent protocol plugin, signed outbound webhooks, and grounded citations. Two of those matter more than the other two. Here is what is actually new versus incremental.
What shipped in the Herald release
Dubbed "The Herald Release", v0.20.0 turns the open-source agent from a text-based assistant into something that speaks, coordinates with other agents over a standard protocol, pushes signed events to external systems, and backs its research with verifiable citations.
The scale of the window is worth noting in concrete terms. Since v0.19.0 on July 20: approximately 3,650 commits, 1,400 merged PRs, 5,200 files changed, 1,200 issues closed, and 650+ contributors. That is roughly two weeks of work. Hermes Agent now sits at over 224,000 GitHub stars.
The four headline features, ranked by practical weight:
- Latency. First token drops from ~4.3 seconds to ~0.9 seconds; the telemetry gate is 54x faster. If you run Hermes in a Slack or chat context, this is the change you will notice most.
- A2A v1.0. A bundled plugin implementing the Agent-to-Agent protocol, closing a feature request opened in 2025.
Hermes can now discover, talk to, and be driven by other A2A-compatible agents.
Voice. Real-time conversation you can interrupt mid-sentence, a custom wake word that works from across the room, and spoken replies inside WhatsApp, DingTalk and LINE.
Grounded citations. Research responses now return with source links on every claim - closer to Perplexity than a bare language model answer.
The voice feature gets the most demo attention. The latency and A2A changes are the ones that compound over time.
The model and the agent are two separate things
This is the part most coverage conflates. Hermes Agent is a runtime - an open-source framework you install, configure, and point at a language model. The Hermes model family (Hermes 3, Hermes 4, Hermes 4.3) is a separate set of weights that Nous Research publishes on HuggingFace.
You do not have to use a Hermes model to run Hermes Agent. Most teams using it in 2026 are not.
Nous Research's documentation describes the Portal as the company's "unified subscription gateway" and the recommended way to run Hermes Agent: a single login provides access to more than 300 models from multiple providers, including Claude, GPT, Gemini, and DeepSeek models, plus a Tool Gateway that routes web search, image generation, text-to-speech, browser automation, and terminal sandbox services through the same subscription.
Under the hood, the Portal routes each model to the backend best suited for it - some models go through OpenRouter, others through proprietary or secondary providers, and the routing for a given model can change over time.
That means your model choice is a config line, not an architecture decision.
You can switch between Claude Sonnet 4.6 for code and Gemini 3 Pro for long context with /model mid-session - no new credentials, no top-ups.
The non-obvious consequence: Nous Research's open-weight model work (Hermes 4.3, Psyche, DeepHermes) feeds its research credibility but is increasingly orthogonal to how the agent runtime is actually used by teams. The agent has become a multi-model harness first.
Beagle in action#eng-ops, 10:45amWhat is genuinely new about Hermes 4.3, the model
If you do want to run a local Hermes model, the one worth knowing about is Hermes 4.3. Released August 25, 2025 and built on ByteDance's Seed 36B base model, Hermes 4.3 was trained using Nous Research's Psyche decentralized training network rather than a traditional centralized GPU cluster.
Hermes 4.3 is Nous Research's first production model post-trained entirely on the Psyche network, which uses the DisTrO optimizer to efficiently communicate between training nodes spread out through data centers over the open internet and secured by the consensus of the Solana blockchain. Whether that architecture survives contact with larger training runs is genuinely unknown - Nous Research acknowledges this is still experimental.
The model numbers are strong for its size. Benchmark scores from the model card: MATH-500 at 93.8%, MMLU at 87.7%, BBH at 86.4%, AIME 24 at 71.9%, GPQA Diamond at 65.5%.
Hermes 4.3 was trained with an extended context length of up to 512K and nearly matches - and in some cases exceeds - the performance of Hermes 4 70B at half the parameter cost.
The post-training corpus is one place where the work is clearly non-incremental. The post-training corpus grew from 1M samples and 1.2B tokens to approximately 5M samples and 60B tokens, blended across reasoning and non-reasoning data.
| Hermes 3 8B | Hermes 4 70B | Hermes 4.3 36B | |
|---|---|---|---|
| Base model | Llama 3.1 8B | Llama 3.1 70B | ByteDance Seed 36B |
| Training method | Centralized | Centralized | Psyche (decentralized) |
| Context window | 128K | 128K | 512K |
| Post-training tokens | ~1.2B | ~1.2B | ~60B |
| MATH-500 | - | - | 93.8% |
| Best fit | Local, low-VRAM | Quality ceiling | Local reasoning work |
One practical catch: the Llama 3.1 base models underlying Hermes 3 use a context window of 128K tokens; if you built pipelines that pass very long context to Hermes 3 and are migrating to Hermes 4.3, verify the effective context length for your use case. The 512K figure is from training, not necessarily from every inference backend.
Where the economics actually sit
Closed models (GPT-4o/5, Claude 4.5, Gemini) maintain a measurable lead on frontier benchmarks in 2026. But open models close 70-90% of the capability gap at 5-10× lower per-token inference cost.
For Hermes Agent specifically, the cost picture depends on which deployment path you take:
Nous Portal (hosted). The free tier is $0/month with $0 in monthly credits, capped at 50 RPM and 500K TPM, limited to the free-model catalog - it is not a trial of the paid product. The Tool Gateway and the paid model catalog both sit behind a subscription. The Plus tier at $20/month is the real entry point, and it returns $22 in credits.
OpenRouter. OpenRouter gives Hermes access to 200-plus models through a single API endpoint and lets you route by cost per token in real time, sending each turn to the cheapest model that clears your quality bar.
Local / Ollama. If privacy is a concern, run Ollama locally with Hermes 3 8B - API cost drops to zero but you need 16GB+ of VRAM.
The Hermes 4.3 GGUFs comfortably sit in the VRAM of off-the-shelf GPUs.
The self-hosting math only works at consistent load. Idle accelerators are expensive. Unpredictable demand is expensive. A team with spiky, business-hours-heavy usage should stay on a hosted API path.
One thing the v0.20.0 release actually improves here: the 54x faster telemetry gate reduces token overhead per session, which compounds across high-volume agent workflows. It is not a pricing change, but it shrinks the effective per-task cost.
Hermes Agent v0.20: common questions
What is new in Hermes Agent v0.20.0?
v0.20.0 ships conversational voice with streaming TTS, barge-in, and on-device wake words; A2A v1.0; signed outbound webhooks; and a grounded-citations skill. The latency improvement - first token from 4.3 seconds to 0.9 seconds - is the change most teams will feel in daily use.
Do I need to use a Hermes model to run Hermes Agent?
No. The agent runtime and the model family are separate. A Nous Portal subscription provides access to more than 300 models from multiple providers, including Claude, GPT, Gemini, and DeepSeek. Most teams in production route to a frontier closed model rather than a Hermes open-weight model.
What is Hermes 4.3 and how does it differ from Hermes 4?
Hermes 4.3 was trained with an extended context length of up to 512K and nearly matches the performance of Hermes 4 70B at half the parameter cost. The key technical difference is training method: it is the first production model post-trained entirely on the Psyche decentralized network, using the DisTrO optimizer across nodes secured by the Solana blockchain.
What does the Nous Portal free tier actually give you?
The free tier is $0/month with $0 in monthly credits, capped at 50 RPM and 500K TPM, and limited to the free-model catalog. It is not a trial of the paid product - the Tool Gateway and the paid model catalog both sit behind a paid subscription.
Is Hermes Agent production-ready for teams using Slack?
The agent maintains memory across sessions and can be used from a terminal or through messaging platforms including Telegram, Discord, Slack, WhatsApp, and Signal.
The v0.20.0 speed improvements make it more usable in real-time channels. The honest caveat: versioned releases are shipping every two to three weeks, so any configuration-sensitive workflow should pin to a specific release tag rather than tracking latest.