Hermes Agent Streaming + Which AI Models Need The New stream_max_lifetime Cap, With Real GoldieBench Scores For Opus 5, GPT-5.6 Sol, Kimi K3, MiniMax M3 And Gemini 3.6 Flash

Hermes agent streaming now has an opt-in wall-clock cap, and the model you run decides whether that cap helps you or cuts off your best answers.

A fast model that finishes quickly can live happily under a tight cap.

A slow reasoning model that thinks for minutes needs the cap off, or set very high.

Below are the GoldieBench scores I read on the live pages on 2 November 2026, and how I'd pair each model with the new setting; if you'd rather plan it with my team, book a free SEO strategy session.

What Changed In Hermes Streaming On 1 November

Commit ac8037c05b added agent.stream_max_lifetime, merged to main on 1 November 2026 and not yet in a tagged release (the latest tag is v0.22.5).

It caps the total time of one streaming response, in seconds, and the default 0 leaves it off.

The older stale detector resets on every chunk, so a stream that keeps trickling data but never ends was never cut.

The cap never resets, never fires before the effective stale-stream timeout, and restarts for each retry attempt.

agent:
  stream_max_lifetime: 480

It's config.yaml only, since follow-up b484be5610 removed the environment variable.

The Model Shortlist, With Real GoldieBench Scores

Ranks come from the GoldieBench leaderboard, which lists 33 models, 50 tasks and 1,310 live demos.

Fusion sits at #1 with 8.59, but it's an OpenRouter multi-model panel rather than one model you'd tune a stream cap around.

What These Scores Mean For A Stream Cap

GoldieBench scores are Julian's 0 to 10 verdicts on one-shot builds: same fixed prompts, one HTML file out.

They measure first-try quality, not stream length or latency.

So read them as "how good is the answer if it gets to finish", and pick the cap from how long that model's healthy replies take on your own work.

A high score is exactly why you don't want a cap cutting a model off mid-answer.

Measure first: ask your agent for the longest reply that finished normally last week, then set the cap well above it.

Check The Transport Before You Pick A Model

Only two stream types are capped: Chat Completions and Anthropic Messages.

Anything on the Responses transport, codex_responses and custom providers declared with it, runs uncapped, and so does Bedrock.

Hermes sets the wire protocol per endpoint, as the configuration docs describe with api_mode: chat_completions, codex_responses or anthropic_messages.

So the same model can be capped through one endpoint and uncapped through another.

Ask your agent which api_mode your main model resolves to before you rely on the cap.

Top Scorers: Claude Opus 5 And GPT-5.6 Sol

Claude Opus 5 has the highest single-model average on the board at 8.27 across 50 tasks, with 13 gold medals.

Its own page lists a trade-off worth noting here: reasoning by default eats token budgets unless tuned per call.

Long reasoning means long streams, and the Hermes docs say to leave the cap at 0 for turns that legitimately stream for a long time.

GPT-5.6 Sol sits at 8.16, and its page says it was benched at medium reasoning effort through OpenRouter.

For both, I'd leave the cap off for research work and set a generous one only on a customer-facing bot.

Long Unattended Runs: Kimi K3 And GLM-5.2

Kimi K3 scores 7.89 across 50 tasks with 8 golds.

Its page says it's wired into the Agent OS as the kimi-k3 Hermes profile for long unattended runs, where a slow-but-right model beats a fast-but-forgetful one.

GLM-5.2 scores 7.77 across 47 tasks, and its page names it the default Agent OS model for long-context work.

Both point the same way: these are jobs where a hard wall-clock cap is the wrong tool.

The stale detector and the HERMES_STREAM_STALE_GIVEUP breaker, five straight stale kills by default, are the guards that fit long runs.

Fast And High-Volume: Gemini 3.6 Flash And MiniMax M3

This is where the cap earns its place.

Gemini 3.6 Flash scores 7.08 across 50 tasks, and its page says it builds full skill-spec 3D games in about 60 to 120 seconds of generation.

A model that usually finishes that fast can take a tight cap without losing good answers.

MiniMax M3 scores 7.97 across 47 tasks, and its page describes it as designed for high-volume agent workloads.

High volume means more chances for one stuck stream to hold a session, so a cap set above its normal reply length is cheap insurance.

Remember the floor: anything below the stale timeout, 180 seconds at the default, behaves as the stale timeout.

Panels: Hermes MoA

Hermes MoA scores 8.17 across 47 tasks.

Its page explains how it works: one prompt goes to a panel of frontier models in parallel, then an aggregator merges the drafts into one answer.

That's several model calls per answer, so think of the cap per call rather than per finished answer.

Each streaming attempt gets its own cap clock, and a cut attempt doesn't count toward the stale breaker, because the provider was still responding.

Local Models: Different Timers Entirely

The GoldieBench local leaderboard ranks seven on-device models, led by Qwable 5 27B Coder at 7.14 across 41 tasks.

Local endpoints get different defaults in Hermes: a socket read timeout raised to 1,800 seconds and a 900-second stale ceiling, because they send no pings during prefill.

Since the cap is floored at the effective stale timeout, a local model at that ceiling won't be cut before 900 seconds.

Tune agent.local_stream_stale_timeout first.

A 27B coder on a laptop can spend a long time reading a big prompt before the first token, and that wait is exactly what the raised ceiling is for.

Want the local setups other members run? Join the AI Profit Boardroom and compare notes on the coaching calls.

My Pairing Rules

• Opus 5, GPT-5.6 Sol, Kimi K3, GLM-5.2 on long work: cap off.
• Gemini 3.6 Flash or MiniMax M3 on a customer bot: cap a few minutes above the longest normal reply.
• Any model on codex_responses or Bedrock: the cap won't apply, so rely on the other timers.
• Local models: tune the 900-second ceiling, then decide on a cap.

Scheduled jobs follow the same logic, because cron runs still stream on the wire; see my Hermes agent cron job model guide for which model to pin per job.

Want the trainings and the Agent OS? Join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe five-step config for live replies and a hard stop on endless streams
• AI Income Deskwhat live replies and the new safety cap mean for a small business bot
• AI Tool Verdicta nine-point scorecard on how Hermes catches stuck streams
• agentos.guidefive copy-paste prompts that set up live replies and a stream cap
• aiprofitboardroom.comthe complete Hermes Agent Streaming guide on the AI Profit Boardroom blog

Pick your model from the table, set the cap only where replies are short, and if you want my team to check the plan, book a free strategy session before you tune Hermes agent streaming.