Model reference · open weights
HARC-Qwen2.5 is an open-weight language model from microsoft. HARC-Qwen2.5-7B-Instruct (BF16) weighs 15.2 GB; the smallest configuration that runs it is 2× RTX 3060 12 GB.
Summary of the microsoft/HARC-Qwen2.5-7B-Instruct model card, 2026-10-01
What it is
| Released by | microsoft |
|---|---|
| Released | 2026-07-02 |
| Parameters | 7.6B |
| VRAM | 15.2 GB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory | |
|---|---|---|---|---|
| 8K each | 32K each | |||
| RTX 3060 12 GB … RTX 4060 Ti 16 GB 2 smaller cards | — | — | — | |
| RTX 3090 24 GB | 15 | 3 | all 32K | 23.4 GB |
| RTX 4090 24 GB | 15 | 3 | all 32K | 23.4 GB |
| RTX 5090 32 GB | 31 | 7 | all 32K | 31.0 GB |
| L40S 48 GB | 59 | 14 | all 32K | 44.0 GB |
| A100 80 GB | 132 | 33 | all 32K | 78.2 GB |
| H100 80 GB | 122 | 30 | all 32K | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | 156 | 39 | all 32K | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | 184 | 46 | all 32K | 107 GB |
| H200 141 GB | 250 | 62 | all 32K | 138 GB |
| B200 180 GB | 329 | 82 | all 32K | 176 GB |
| 2× RTX 3060 12 GB tensor parallel | 13 | 3 | all 32K | 11.6 GB a card |
| 2× RTX 4060 Ti 16 GB tensor parallel | 29 | 7 | all 32K | 15.4 GB a card |
| Requests at once | 8K tokens each | 32K tokens each |
|---|---|---|
| 1 | 16.5 GB | 17.9 GB |
| 5 | 18.4 GB | 25.4 GB |
| 8 | 19.8 GB | 31.1 GB |
| 16 | 23.5 GB | 46.1 GB |
| 32 | 31.1 GB | 76.2 GB |
| 64 | 46.1 GB | 136 GB |
One card, with vLLM's small-card settings.
From the model card
HARC safety-alignment LoRA merged into Qwen/Qwen2.5-7B-Instruct (full standalone model).
Part of the HARC release; see paper arXiv:2607.00572.
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("microsoft/HARC-Qwen2.5-7B-Instruct", torch_dtype="bfloat16", device_map="auto")
tok = AutoTokenizer.from_pretrained("microsoft/HARC-Qwen2.5-7B-Instruct")
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.