Hermes 4 70B

nous-research/hermes-4-70b · by Nous Research · decoder-only-transformer · Built on meta/llama-3.1-70b

Pricing — 1 offering(s)

OpenRouter

About this offering: Hermes 4 70B (Nous Research, released 2025-08-26): a hybrid-reasoning fine-tune of Meta's Llama 3.1 70B with a toggleable thinking mode and a 131K-token context, under Meta's Llama community license rather than Apache or MIT. Retired: OpenRouter listed no endpoints for it by 2026-09-28, so the prices shown are the last published rates.

Input tokens

Retired — no longer sold · last price $0.13 / 1M tokens (input), until 2026-09-27

$0.13 / 1M tokens (input) Historical 2025-08-26 → 2026-09-27

Priced through OpenRouter: Nous Research has no first-party paid API for this model, and OpenRouter forwarded…

Priced through OpenRouter: Nous Research has no first-party paid API for this model, and OpenRouter forwarded every request to a single upstream provider.

Output tokens

Retired — no longer sold · last price $0.40 / 1M tokens (output), until 2026-09-27

$0.40 / 1M tokens (output) Historical 2025-08-26 → 2026-09-27
Same sourcing caveat as the input tier.

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Hermes 4 70B fits into a cost-aware routing setup

Capability profile

How Hermes 4 70B rates across core capability dimensions, with the task-level evidence behind each rating.

reasoning Strong

Nous's own Hermes 4 Technical Report (reasoning mode): MATH-500 95.6, AIME'24 73.5, AIME'25 67.4, MMLU-Pro 80.7, BBH 87.8. Strong for a 70B open model; the parenthesised non-reasoning-mode figures in the report are materially lower, so the toggle matters.

coding Moderate

LiveCodeBench v6 (Aug-2024+) ~50.5 in the technical report — solid but not a coding specialist.

instruction following Strong

The Hermes line is positioned around steerability and adherence to system prompts / structured-output schemas with minimal refusal behaviour (RefusalBench 59.5, Arena-Hard v1 90.1 in the report).

tool use Moderate

Structured outputs and function calling are a stated focus; no independent agentic-benchmark score published.

context window Moderate

131,072 tokens — inherited from the Llama 3.1 70B base.

multilingual Moderate

Inherits Llama 3.1's multilingual coverage; not separately optimised.

speed Moderate

Dense 70B; reasoning mode adds latency proportional to trace length.

cost efficiency Strong

$0.13 / 1M input, $0.40 / 1M output via OpenRouter — cheap for a 70B model with this reasoning quality, and open weights for self-hosting.

Operator guidance

Route here for open-weights reasoning with high steerability at low cost — the Hermes line's traditional niche. For frontier reasoning accuracy, a larger open model (Qwen 3 235B, DeepSeek V4) or a proprietary reasoning model will outperform it. The Meta Llama 3 Community License (below) is more restrictive than the Apache-2.0 of most other open models in this registry — check it if licensing is a constraint.

Use cases

Open-weights reasoning at 70B scale where steerability and low refusal rates matter (agent backends, structured extraction, roleplay/persona work)
Self-hosted deployment where a togglable reasoning mode is wanted without running a frontier-size model
Tasks needing strong math/logic at open-model cost

Limitations

Meta Llama 3 Community License — a derivative-model license with use restrictions, not a fully-open Apache/MIT license
Benchmark figures are from Nous's own technical report; no independent third-party leaderboard reproduction found
No first-party hosted API — served here via OpenRouter (single upstream provider)
Reasoning-mode benchmarks are substantially higher than non-reasoning mode; a fair comparison depends on which mode the caller uses