THE CASCADE · hermes moa serve
mode: cascade · preset: cascade-rlm-conv · pinned always-on service

GPT-5.5-class answers in about 4 seconds.

Don't predict which questions are hard — observe agreement between fast agents, and only pay for intelligence when they disagree.

AIME 2025 accuracy
97%
58 / 60 — GPT-5.5 solo scores 98% but takes 40s to get there
Median latency
3.9s
vs 40s for GPT-5.5 solo — a ~10× wall-clock difference
Marginal cost / query
~$0.003
frontier calls ride a ChatGPT plan at $0 — touched on 17% of queries
01 How it works

Agreement is the router.

No difficulty classifier, no learned router for this preset. Four cheap voters answer in parallel; if they agree, that's the answer. If they don't, and only then, a frontier model is paid for.

Cascade tier-flow diagram A query fans out to four Tier 0 voters on Cerebras wafer inference. If three or more agree, the answer returns immediately. On disagreement, GPT-5.5 re-solves the problem from scratch in Tier 1 and that answer is returned instead. QUERY user prompt TIER 0 — PARALLEL VOTERS Cerebras wafer inference, ≤6 rounds, sandboxed exec gpt-oss-120b single-pass · voter A gpt-oss-120b single-pass · voter B gemma-4-31b RLM agent · voter C reason → run_python → observe gemma-4-31b RLM agent · voter D reason → run_python → observe ≥3/4 agree? ≥3/4 agree · 83% RETURN ANSWER median 3.0s · 98% precision 83% of AIME traffic disagree · 17% TIER 1 — CLEAN ARBITRATION GPT-5.5 (ChatGPT-plan OAuth) re-solves FROM SCRATCH voters' outputs are deliberately discarded showing the voters' work anchors the arbiter — measured 7/9 vs 98% solo when contaminated so it never sees them RETURN ANSWER 3.9s median, overall 17% of AIME traffic

Tier 0 runs four voters in parallel on Cerebras wafer inference: two gpt-oss-120b single-pass calls plus two gemma-4-31b agents running an RLM loop (reason → run_python → observe, capped at 6 rounds, sandboxed execution). Extracted answers are normalized and compared. Tier 1 only fires on disagreement — GPT-5.5 re-solves the problem from scratch, never seeing what the voters produced.

02 Measured results

Where the seconds and the accuracy actually go.

Every number on this page comes from timed runs, not estimates. Sample sizes are small (see honest limits) — treat single-point comparisons as directional.

Accuracy vs median latency, AIME 2025 Scatter plot of AIME 2025 accuracy against median latency in seconds on a log scale. Cascade-rlm-conv reaches 97% accuracy at 3.9 seconds, close to GPT-5.5 solo's 98% at 40 seconds, at roughly one tenth the latency. 50% 60% 70% 80% 90% 100% 1s 2s 5s 10s 20s 40s MEDIAN LATENCY (s, log scale) AIME 2025 ACCURACY gemma-4-31b solo 68% · 1.0s gpt-oss-120b cere-moa 90% · 19s cascade-wafer4 93% · 4.4s cascade-rlm 92% · 3.6s Fable solo 100% · 24s GPT-5.5 solo 98% · 40s CASCADE-RLM-CONV 97% · 3.9s (shipped)
shipped config (cascade-rlm-conv) other configurations / solo models

Accuracy vs. median latency, AIME 2025, latency on a log axis. The shipped config sits closest to the top-left corner of any config that isn't a raw frontier or Fable solo call.

Share of queries that touch the frontier model Bar chart of the fraction of queries reaching the frontier model. GPT-5.5 solo is 100% by definition. cascade-rlm-conv touches the frontier on 17% of queries, consensus-or-frontier 15%, cascade-wafer4 5%, cascade-rlm 1.7%. 0% 25% 50% 75% 100% GPT-5.5 solo 100% cascade-rlm-conv 17% shipped config consensus-or-frontier 15% cascade-wafer4 5% cascade-rlm 1.7%

Fraction of queries that reach the frontier model. cascade-rlm-conv spends more frontier calls than the plainer wafer-only cascades in exchange for closing most of the remaining accuracy gap — still a 6× reduction from calling GPT-5.5 on every query.

The RLM lift for gemma-4-31b Grouped bar chart comparing gemma-4-31b solo against gemma-4-31b as an RLM agent. AIME 2025 accuracy rises from 68% to 82%. GPQA-numeric, n=25, rises from 40% to 56%. solo RLM agent 0% 20% 40% 60% 80% 100% 68% 82% AIME 2025 40% 56% GPQA-numeric (n=25)

gemma-4-31b gains 12–16 points from running as an RLM agent (reason → run_python → observe) instead of answering solo. gpt-oss-120b shows no such benefit — the working rule so far: externalized reasoning helps models that don't already reason internally, and doesn't help ones that do.

cross-set check HMMT Feb 2025 — 90% @ 5.5s clean arbiter went 7/7 on the disagreement tail — the anchoring fix generalized past AIME.
03 How to configure it

The exact preset, verbatim.

cascade-rlm-conv is a preset shipped with hermes moa serve — nothing here is bespoke glue code.

# preset: cascade-rlm-conv
mode: cascade
cascade:
  min_consensus: 3
  clean_arbiter: true
reference_models:
  - provider: custom:cerebras
    model: gpt-oss-120b
  - provider: custom:cerebras
    model: gpt-oss-120b
  - provider: custom:cerebras
    model: gemma-4-31b
    agent: rlm
  - provider: custom:cerebras
    model: gemma-4-31b
    agent: rlm
aggregator:
  provider: openai-codex
  model: gpt-5.5
reference_max_tokens: 8000
reference_quorum_grace: 1.0
# start the server
hermes moa serve --port 8655
# any OpenAI-compatible client works — this one's Python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8655/v1", api_key="not-needed")

resp = client.chat.completions.create(
    model="moa:cascade-rlm-conv",
    messages=[{"role": "user", "content": "..."}],
)

The command above is the general-purpose invocation — pick any port for your own deployment. The pinned production instance for this preset runs as the cascade endpoint (localhost :8655 on the host). Any OpenAI-compatible client or agent harness can point at it, including client-side tool calls — they're forwarded straight through to the underlying models.

04 Honest limits

Where this doesn't apply.

⚠ read before you route production traffic here

  • Competition-math-class evaluation only. AIME, HMMT, GPQA-numeric — exact-answer domains where a consensus vote can be checked by string/value normalization. This is not evidence for open-ended or prose tasks.
  • ~92–95% is the wafer-only ceiling. The remaining errors are systematic, not random: all four Tier 0 voters can agree on the same wrong answer, and no consensus rule catches a unanimous mistake. That failure mode is exactly why the frontier arbiter tier exists.
  • Prose and code lanes route elsewhere. Advisory context from other agents measurably hurts code-editing accuracy in our tests — the cascade pattern here is scoped to exact-answer math, not applied uniformly.
  • Small samples. AIME 2025 is n=60; the GPQA-numeric RLM-lift comparison is n=25; several other evals cited here are n=20. Treat single-point deltas as directional, not final.
Post-cutoff revalidation (2026-07-05) — the architecture got STRONGER. On AIME 2026 and HMMT Feb 2026 (administered after every model's training cutoff), the frontier solos dropped (GPT-5.5: 97% / 85%; Fable: 97% / 90%) while the cascade held: cascade-rlm scored 97% on AIME 2026 — tying both frontier models — and 95% on HMMT Feb 2026, beating GPT-5.5 by 10 points and Fable by 5, at 3.9–5.7 s median (they take 22–38 s), touching the frontier on 3–20% of problems. On contamination-proof problems, consensus-of-fast-agents now BEATS the frontier solos it escalates to. The default preset is now cascade-rlm; numbers above from the pre-2026 sets remain as measured (cascade numbers revalidated; frontier-solo numbers there are contamination-flattered).