Hermes 3 Llama 3.1 8B
NousResearch's Hermes fine-tune of Llama 3.1 8B. Stronger system-prompt adherence, JSON output, role-play, and agent steering than the base Llama.
Performance evidence
Performance statements in this entry have not been tied to a published, configuration-matched run. Treat speed figures as unverified estimates, not validated measurements. Compare the full configuration of a published benchmark before relying on throughput or latency.
Positioning
Hermes 3 is the uncensored / less-aligned alternative on the Llama 3.1 8B base. Right pick for security research, red-team work, technical writing on dual-use topics, or any case where the base Llama's refusal layer gets in the way of legitimate work.
Strengths
- Refusals dramatically reduced vs base Llama 3.1 8B without losing instruction quality.
- Same VRAM, same Llama license — drop-in replacement.
- Tool-use compatibility preserved.
Limitations
- Niche use case — most users don't need this; default to Llama 3.1 8B.
- Slightly weaker on creative writing than base Llama (alignment training adds polish).
- Reduced refusals can be too eager — produces content that requires judgment to use.
Historical performance claims (unverified) on RTX 4090
- Q4_K_M (4.6 GB): 90–110 tok/s decode
- Q5_K_M (5.6 GB): 80–95 tok/s
- Q8_0 (8.5 GB): 65–80 tok/s
Should you run this locally?
Yes, for security/research work where base Llama's refusals are blocking legitimate tasks. No, for general chat — the base Llama 3.1 8B is the right default.
How it compares
- vs Llama 3.1 8B (base) → Hermes 3 is base Llama minus alignment layer. Pick base for general use, Hermes for technical/research work.
- vs Hermes 3 Llama 3.1 70B → 70B is meaningfully smarter at higher VRAM cost.
- vs Dolphin 3.0 Mistral 24B → similar philosophy, different base model. Dolphin is bigger and on Apache base.
Run this yourself
ollama pull nous-hermes-3:8b-llama-3.1-q4_K_M
ollama run nous-hermes-3:8b-llama-3.1-q4_K_M
›Why this rating
7.7/10 — the right pick when Llama 3.1 8B's alignment refusals get in the way. NousResearch's Hermes 3 strips the over-cautious layer while keeping instruction-following intact. Loses points only on niche use case.
Overview
NousResearch's Hermes fine-tune of Llama 3.1 8B. Stronger system-prompt adherence, JSON output, role-play, and agent steering than the base Llama.
Family & lineage
How this model relates to others in its lineage. Family members share architecture and training-data roots; parent / children edges record direct distillation or fine-tune relationships.
Strengths
- Excellent system-prompt obedience
- JSON / structured output
- Agent-friendly
Weaknesses
- Inherits Llama 3.1 license
Reviewed quality benchmarks
First-party rows were run by RunLocalAI; reviewed community rows are labeled in the data. Every row links to the raw test-run log.
| Benchmark | Quant | Runtime / Hardware | Score | Raw log |
|---|---|---|---|---|
HumanEval+ tested 2026-07-20 | Q4_K_M | ollama-0.32.1-vast5090 rtx-5090 | 41.5/100 | Gist → |
Q4_K_M note:First-party HumanEval+ on a rented Vast.ai RTX 5090 32GB instance via Ollama 0.32.1. Generation via evalplus_openai_generate.py, scoring via evalplus.evaluate (native Linux, no Windows shim needed). No DB credentials were placed on the rented host — results were pulled back and ingested from the operator's machine.
Want to verify? Every row links to its Gist with full stdout and stderr of the run. The runner script is in the public repo (scripts/run-humaneval-plus.ts) — reproducible end-to-end. Browse all coding scores at /benchmarks/coding.
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| Q4_K_M | 4.9 GB | 6 GB |
| Q8_0 | 8.5 GB | 10 GB |
Get the model
Ollama
One-line install
ollama run hermes3:8bRead our Ollama review →HuggingFace
Original weights
Source repository — direct quantization required.
Benchmarks
Real measurements on real hardware. Numbers ship with the runner version, quant, and date.
| Hardware | Provenance | Quant | Ctx | Tokens / sec | TTFT | Date |
|---|---|---|---|---|---|---|
| NVIDIA GeForce RTX 3080 16GB (Mobile) | EditorialM | Q4_K_M | 4K | 81.5tok/s | 357 ms | Jun 2, 26 |
What to do next
Got this model running on real hardware? Share what you measured — the form arrives with the model pre-selected.
Hardware that runs this
Cards with enough VRAM for at least one quantization of Hermes 3 Llama 3.1 8B.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run Hermes 3 Llama 3.1 8B?
Can I use Hermes 3 Llama 3.1 8B commercially?
What's the context length of Hermes 3 Llama 3.1 8B?
How do I install Hermes 3 Llama 3.1 8B with Ollama?
Source: huggingface.co/NousResearch/Hermes-3-Llama-3.1-8B
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Hermes 3 Llama 3.1 8B runs on your specific hardware before committing money.