Hermes Compaction + Which AI Models To Run As Main And Summary Model, With Real GoldieBench Scores For Opus 5, Kimi K3, MiniMax M3 And Gemini 3.6 Flash

Hermes compaction runs on two models, not one: your main chat model decides when the conversation is full, and a separate summary model writes the handoff that replaces the middle.

Pick the pair badly and the summary call fails, which the Hermes docs name as the most common cause of poor compaction.

Below are the GoldieBench scores I read on the live leaderboard and models pages on 3 November 2026, and how I would pair them for automatic compaction, the new prepare_ahead option and micro-compaction.

If you would rather plan the stack with my team, book a free SEO strategy session.

Why Hermes Compaction Needs Two Models

The configuration reference says the compression summariser is a separate LLM call that you can point at any provider or endpoint.

It lives under auxiliary.compression, and an empty model field means Hermes reuses your main chat model:

auxiliary:
  compression:
    model: ""                                       # Empty = use main chat model. Override with e.g. "google/gemini-3-flash-preview" for cheaper/faster compression.
    provider: "auto"                                # Provider: "auto", "openrouter", "nous", "codex", "main", etc.
    base_url: null                                  # Custom OpenAI-compatible endpoint (overrides provider)

The trigger is always worked out from the main model's window: threshold of 0.50 means half of that window.

So the main model sets the timing, and the summary model sets the quality of what survives.

The One Rule Every Pairing Must Pass

The developer guide is blunt: the summary model needs a context window at least as large as the main model's.

The whole middle of the conversation goes out in one call, so a smaller window returns an error and the middle is dropped with no summary.

That rule matters more than any benchmark score on this page.

A top-scoring model with a smaller window is the wrong summariser for a 1M-context main model.

The Shortlist, With Real GoldieBench Scores

Scores, task counts and ranks come from the GoldieBench leaderboard; the notes come from the models page.

One honest caveat: GoldieBench scores one-shot builds of games, visuals, sims and pages, not summaries.

Read the numbers as general capability, then let the window rule above decide the pairing.

Main Model Picks For Long Sessions

For the chat model that decides when to compact, the window drives how often you pay for a summary.

MiniMax M3 lists a 1,048,576-token window on the leaderboard, and Kimi K3 and GLM-5.2 are both described as 1M context on the models page.

On windows that big, the default 50% trigger fires late, so the docs offer threshold_tokens: their example is 256000, to compact a 1M model at 256K instead of 500K.

Claude Opus 5, at 8.27 the highest-scoring single model on this list, is the pick when quality per turn matters more than how long a session runs.

If you swap between big and small windows, use model_thresholds so each one compacts at its own point:

compression:
  threshold: 0.50
  model_thresholds:
    "glm-5.2": 0.40
    "glm-5.2-1M": 0.25
    "claude-sonnet": 0.35
    "openai-codex:astra": 0.85   # only on the Codex OAuth route (272K cap)

Summary Model Picks For Automatic Compaction

Batch compaction makes one larger call per compaction, so quality beats speed here.

The simplest safe choice is to leave auxiliary.compression.model empty, which reuses the main model and passes the window rule by definition.

If you want a separate summariser, match windows first: a 1M-context main model such as Kimi K3 wants a 1M-class summariser such as MiniMax M3 or GLM-5.2.

The docs' own provider example looks like this:

auxiliary:
  compression:
    provider: nous
    model: gemini-3-flash

What prepare_ahead Changes For Your Model Choice

The v0.22.6 release on 2 November 2026 added opt-in prepared compaction, from commit 32a636bc8d.

compression:
  prepare_ahead: false                              # Opt-in: prepare the next summary in the background — see below

Flip it to true and the next summary gets drafted early, in the background, as soon as the prompt sits within 12% below the trigger; Hermes then swaps it in at compaction if those same messages are still there.

That hides the summary model's latency, which makes a slower, higher-scoring summariser easier to live with.

The cost is one auxiliary call that may go unused, and it never applies to manual /compress, plugin context engines or native compaction.

Micro-Compaction Wants A Small, Fast Model

The micro-compaction guide flips the advice.

It runs a summary pass after every finished turn, so latency is felt each time, and the guide says reasoning models are a poor fit for this mechanical job.

It quotes one setup where a local 7B 4-bit instruct model took about 31 seconds per pass.

On GoldieBench, Gemini 3.6 Flash is described as faster with fewer tokens, which is the right shape, and Kimi K2.7 No-Think is listed as a no chain-of-thought mode, though that variant has no scored verdicts yet.

compression:
  micro_compact: true             # default: false
  micro_compact_every_n_turns: 1  # cadence — how often a pass runs
  micro_compact_defrag_threshold_tokens: 2000

Micro-compaction is off by default, and every pass breaks the provider prompt cache, so test it before you commit.

Three Pairings I Would Start With

Before you pair a separate summariser with any main model, confirm its window covers the main model's on the provider you use.

Plugin Engines Take The Model Choice Away

If you set context.engine to a plugin, the context engine docs say it replaces the built-in summariser, and prepare_ahead and the proactive prune stop applying.

context:
  engine: "lcm"          # must match the plugin's name

At that point the engine's own documentation decides which models it calls.

For the stream timers that matter when a slow summariser is streaming, read our Hermes agent streaming model guide.

How I Read These Scores For A Summary Job

A summariser has a narrow brief: keep the goal, the decisions, the file paths and the next steps, and drop the noise.

Hermes hands it a fixed template, so the model does not have to invent a structure, only fill one in accurately.

That is why I treat a gap of a few tenths on the leaderboard as a tie for this job, and let the context window and speed decide.

Where the gap is bigger, such as Opus 5 at 8.27 against Gemini 3.6 Flash at 7.08, I lean on the higher score for batch summaries and the faster model for per-turn passes.

Rank also moves as new models land, so re-check the leaderboard before you change a working pairing.

Members of the AI Profit Boardroom get the Agent OS with these pairings set up, plus coaching calls to test their own.

How To Update And Test

Run hermes update on the CLI; Desktop installer builds get v0.22.6 with the next bundled release.

Then run one long session per pairing and watch the CLI pressure bar climb toward the trigger before the first compaction lands.

Check the summary it writes: if the goal, the files and the next steps are all there, the pairing works.

For the full trainings and the Agent OS that runs these pairings, join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe eight compaction settings I set before a long Hermes session
• AI Income Deskwhether compaction is worth it for shops, service firms and agencies
• AI Tool Verdicta nine-point scorecard for every part of Hermes compaction
• agentos.guidefive prompts that let your agent tune its own compaction
• aiprofitboardroom.comthe complete Hermes compaction guide on the AI Profit Boardroom blog

Match the windows first, then use the scores to break ties, and if you want my team to check your pairing, book a free SEO strategy session before you lock in your models for Hermes compaction.