Model reference · open weights

HarmBench-Llama-2-cls

LLMs cais Text gen 1 build Open weights 138k dl/mo

HarmBench-Llama-2-cls is an open-weight language model from cais. HarmBench-Llama-2-13b-cls (BF16) weighs 26.0 GB; the smallest configuration that runs it is 2× RTX 4060 Ti 16 GB.

For self-hosting, compare this build's requirements below with the GPU you plan to rent.

  • HarmBench-Llama-2-cls is a 13.0B parameter text-generation model developed by cais to serve as the official classifier for text behaviors in the HarmBench framework.
  • It supports both standard and contextual behaviors, operating with a context length of 2048 tokens and released under the MIT licence.
  • The model achieves an average agreement rate of 93.19% with human judgments on a manually labeled validation set, performing comparably to GPT-4.

Summary of the cais/HarmBench-Llama-2-13b-cls model card, 2026-10-01

What it is

Released bycais
Released2024-02-03
Parameters13.0B
VRAM26.0 GB for the weights

What it runs on

Memory and cards for HarmBench-Llama-2-13b-cls (BF16)

26.0 GBweights, file size
819 MBcache per 1K tokens
659 MBruntime overhead, at least
2,048 tokenscontext max
CardRequests at onceContext maxMemory
2K each
RTX 3060 12 GB … RTX 4090 24 GB
4 smaller cards
RTX 5090 32 GB2all 2K31.0 GB
L40S 48 GB10all 2K44.0 GB
A100 80 GB30all 2K78.2 GB
H100 80 GB28all 2K78.1 GB
RTX PRO 6000 Blackwell 96 GB38all 2K93.8 GB
DGX Spark (GB10) 128 GB unified46all 2K107 GB
H200 141 GB64all 2K138 GB
B200 180 GB86all 2K176 GB
2× RTX 4060 Ti 16 GB
tensor parallel
2all 2K15.4 GB a card
2× RTX 4090 24 GB
tensor parallel
11all 2K23.4 GB a card
2× RTX 3090 24 GB
tensor parallel
11all 2K23.4 GB a card
2× RTX 5090 32 GB
tensor parallel
20all 2K31.0 GB a card
Memory needed at each load
Requests at once2K tokens each
128.4 GB
535.1 GB
840.1 GB
1653.5 GB
3280.4 GB
64134 GB

One card, with vLLM's small-card settings.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms