Hermes Agent ships v0.5.0 with 400+ Portal models and Exa support

• Hermes Agent v0.5.0 is positioned as a "hardening release," and the core user-facing expansion is broader model access: Teknium's release thread says Nous Portal now serves "400+ models," while the linked release page adds Hugging Face as a first-class inference provider.
• The provider layer widened again within hours of the release: Teknium's Exa post says Exa is now an official search and scrape backend alongside existing web providers, extending Hermes' terminal-first agent workflow without changing its interface.
• Under the hood, the release page says Hermes swapped in a native Modal SDK backend, activated plugin lifecycle hooks, and added GPT tool-use guidance to stop models from "describing intended actions" instead of making tool calls; Teknium separately said the update should make GPT/Codex models "not so lazy anymore" in a follow-up post.
• Reliability and security work is a major part of the release: according to the release notes screenshot, Hermes removed a compromised LiteLLM dependency, pinned versions, added CI scanning, and fixed Anthropic output limits by moving from a hardcoded 16K cap to model-native limits.

Hermes Agent's biggest practical change is its provider surface. Teknium's release thread says the Nous Portal model provider now serves more than 400 models, and the linked release page describes Hugging Face as a "first-class inference provider" with HF Inference API support, a model picker, live /models probing, and setup-wizard integration. That turns v0.5.0 into more of a routing and access expansion than a workflow redesign.

The same release also adds workflow-facing integrations around that provider layer. The release page says Telegram Private Chat Topics now support project-based conversations with skill binding per topic, and that Hermes' backend replaced the prior swe-rex path with the native Modal SDK using Sandbox.create.aio and exec.aio. Those are concrete plumbing changes for teams running Hermes across chat surfaces or remote sandboxes.

Post-release, Teknium said Exa is now an official search and scrape provider in Hermes. The merged change shown in the Exa integration post adds Exa as a fourth web backend alongside Parallel, Firecrawl, and Tavily, with exa-py added as a dependency and config support via EXA_API_KEY plus web.backend: exa.

Nous and Teknium both framed v0.5.0 as optimization and cleanup, and the most concrete examples are in tool execution and runtime reliability. The release page says Hermes now fires pre_llm_call, post_llm_call, on_session_start, and on_session_end hooks, which gives plugin authors stable lifecycle entry points that were previously missing. The same notes say Hermes added GPT_TOOL_USE_GUIDANCE to keep GPT-family models from narrating actions instead of actually calling tools.

That behavior fix was important enough for Teknium to summarize it separately: in his earlier post, he said Hermes had just absorbed learnings so users' "gpt/codex models" were "not so lazy anymore." In practice, that lines up with the release note language about enforcing tool use rather than accepting descriptive non-actions.

The release also fixes deployment and API edge cases that matter in production. The release notes screenshot says Hermes now supports a full uv2nix build and NixOS module, removed a compromised LiteLLM dependency, pinned versions, regenerated uv.lock with hashes, and added CI scanning as part of a supply-chain audit. On the model side, the same notes say Hermes replaced a hardcoded 16K max_tokens setting on direct Anthropic API calls with per-model native limits—128K for Opus 4.6 and 64K for Sonnet 4.6—addressing "Response truncated" failures and thinking-budget exhaustion.