How to Run Arete-27B Locally: Hardware Needs and Benchmarks

Nous Research released Arete-27B, a 27-billion-parameter Apache 2.0 supervised fine-tune of Alibaba's Qwen3.8-27B trained on roughly 52.8 million tokens at 64k context over two epochs (1,202 steps) to reduce refusals and CCP-framed answers on politically sensitive topics. On the team's own propaganda evaluation, Arete-27B's fact score rose from 2.32 to 2.98 out of 5 and the share of answers containing propaganda fell from 22% to 0.9%, with the freedom-of-expression category moving from 0.89 to 2.44 and a propaganda rate drop from 60% to 1%. General capability held: Hermes Bench scores were within 0.2 percentage points of the base model and Terminal-Bench-2.1 rose from 67% to 70%, while local deployment requires roughly the same VRAM as any dense 27B model depending on quantization.

How to Run Arete-27B Locally: Hardware Needs and Benchmarks Arete-27B is a 27B Apache 2.0 fine-tune of Qwen3.8-27B. Here's the hardware it needs locally and how its benchmarks stack up. What is Arete-27B? Arete-27B is a 27 billion parameter open-weight model from Nous Research, built by post-training Alibaba’s Qwen3.8-27B with supervised fine-tuning SFT only, no reinforcement learning, no preference optimization, no architecture changes. The goal was narrow and explicit: reduce refusals, deflections, and CCP-framed answers on politically sensitive topics that the researchers observed in the base Qwen3.8-27B model, while keeping general capability intact. Running it locally means dealing with a standard dense 27B model, so the hardware math is the same as any other model in that parameter range, with VRAM being the main constraint. TL;DR - Arete-27B is a supervised fine-tune of Qwen3.8-27B, trained on roughly 52.8 million tokens at 64k context over two epochs 1,202 total steps . - The fine-tuning targeted seven value categories including freedom of expression, state ideology, democracy, data privacy, and environmental policy, using an open model called Inkling as the teacher. - On the team’s own propaganda evaluation, Arete-27B’s fact score rose from 2.32 to 2.98 out of 5 , and the share of answers containing propaganda fell from 22% to 0.9%. - The biggest single-category shift was in freedom of expression , where the fact score jumped from 0.89 to 2.44 and the propaganda rate dropped from 60% to 1%. - General capability held steady: Hermes Bench scores for Arete-27B and the base model were within 0.2 percentage points, inside noise, and Terminal-Bench-2.1 moved from 67% to 70%. - As a 27B dense model , running it locally needs roughly the same VRAM footprint as any other model of that size https://goremy.ai/blog/articles/hardware-requirements-for-local-llm-inference-2026-guide , scaling with quantization choice. - The model and its evaluation dataset Propaganda Eval , 803 prompts are both public, letting anyone reproduce or extend the measurement methodology. demands of enterprises. Run it on our own servers. ✓ SOC 2 and HIPAA from day one. ✓ Log every action. ✓ Only the right people get in. ✓ How was Arete-27B trained? Nous Research started from Qwen3.8-27B and applied supervised fine-tuning exclusively, no RL or DPO-style preference tuning was involved. Training ran at 64k context length on about 52.8 million tokens, without sequence packing, across two epochs totaling 1,202 steps 601 per epoch . The optimizer was AdamW with betas of 0.9, 0.98 , weight decay of 0.1, and gradient clipping at 1.0. Learning rate followed a linear warmup from 1e-6 to 1e-5 over the first 120 steps, then cosine decay back down to 1e-6 by the final step. Training used tensor parallelism of 4, bf16 compute with fp32 master weights, and all language-model weights were trainable while the vision encoder stayed frozen the training data was text-only . The multi-token-prediction head was disabled during training. Training loss dropped quickly through warmup, flattened near 1.0 by the end of epoch 1, then stepped down to about 0.8 at the start of epoch 2 as the model re-saw the same data. Gradient norm started around 32 and settled near 1.4 once warmup finished. The evaluated checkpoint is step 1,202. What data and teacher model were used? The post-training data came from a teacher model called Inkling, chosen after benchmarking and qualitative review as the best available open model whose answers on these topics the Nous team was willing to train on. Using an open teacher sidestepped terms-of-use restrictions tied to distilling from certain closed models. The team generated 20,000 prompts spread across seven target categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. A deduplication pipeline checked these against the held-out evaluation set and removed any overlap, then collected both the teacher’s responses and the base model’s own responses for comparison. Every batch of generated data went through pull-request review before being used, mirroring a code-review workflow rather than a one-shot data dump. How does Arete-27B perform on propaganda and factual recall? Nous built a dedicated evaluation, Propaganda Eval , with 803 prompts across the seven value categories, sourced from English-language government statutes, news, and research reports, capped at no more than 13 prompts from any single source. For each prompt, researchers extracted 2 to 4 required facts from a dated source article along with the official CCP framing on that topic. An LLM judge Grok 4.7 scored each model’s answer against those facts and a fixed rubric, without knowing which model produced the answer. Two metrics came out of this: a fact score from 0 to 5 measuring how many required facts were stated, missed, or opposed, and a propaganda flag capturing whether an answer replaced facts with the official state line or endorsed state conduct while stating the facts. Results at temperature 0 with thinking off: | Model | Fact mean 0-5 | Any propaganda | |---|---|---| | Inkling teacher | 3.96 | 0.3% | | Mistral Large 4 | 3.65 | 0.3% | | Nemotron 3 Ultra | 3.22 | 0.6% | | GLM 5.3 | 3.10 | 15% | | Muse Glimmer 30B | 3.02 | 0.5% | | Arete-27B | 2.98 | 0.9% | | DeepSeek V4.1 Flash | 2.39 | 16% | | Qwen3.8-27B base | 2.32 | 22% | mission-critical enterprise applications. Dr. A. Okafor Cardiology · St. Mercy Hospital - Patient records ✓ Granted - Surgery schedule ✓ Granted - Billing export ✕ Denied At temperature 0.6, Arete-27B scored 2.88 with 0.9%-range propaganda behavior, versus 2.24 for the base model. The number of answers scoring a flat 0 no facts stated, or any fact opposed fell from 359 to 216 out of 803 prompts. Every one of the seven value categories improved, with freedom of expression showing the largest swing: fact score up from 0.89 to 2.44, propaganda rate down from 60% to 1%. Nous is explicit that this doesn’t claim to fix every bias in the base model, only to address what they identify as a contributing cause through targeted fine-tuning. Did fine-tuning hurt general capability? Apparently not, based on the two benchmarks reported. On Hermes Bench, an internal general-capability benchmark, Qwen3.8-27B scored 65.77% and Arete-27B scored 65.60%, a gap the team calls noise. On Terminal-Bench-2.1, an agentic coding/terminal benchmark, Arete-27B scored 70% against the base model’s 67%, measured with a like-for-like inference configuration on the same hardware. Nous notes that differences from other published Terminal-Bench-2.1 numbers could stem from inference harness or configuration differences, and they don’t claim the apparent gain as a verified improvement. The practical takeaway: on the evaluations run, value-alignment fine-tuning didn’t cost meaningful capability, at least on these two benchmarks. What hardware do you need to run Arete-27B locally? Arete-27B is architecturally a standard dense 27B parameter language model since the fine-tune touched only the language-model weights the vision encoder was frozen and unused in training . That means VRAM requirements track with any other 27B model, scaling primarily with the quantization format you choose: - Full precision bf16/fp16 : a 27B model at 16-bit weights needs on the order of 54GB of VRAM just to load, before accounting for KV cache and context. This typically means multi-GPU setups or high-end workstation cards. - 8-bit quantization : roughly halves that to around 27GB, putting it within reach of a single high-VRAM consumer or prosumer GPU. - 4-bit quantization : cuts memory further, to somewhere in the 14 to 16GB range depending on the exact quantization scheme, making it plausible on a single mainstream GPU. The model was trained and evaluated at 64k context, so if you intend to use long context locally, budget extra VRAM for the KV cache on top of the weight footprint; cache size grows with context length and batch size. CPU-only or CPU-offloaded inference is possible but will be substantially slower than GPU inference at this parameter scale. As with any open-weight release, check the specific quantized build you download GGUF, AWQ, GPTQ, etc. for its exact memory footprint, since different quantization methods pack weights slightly differently. Is Arete-27B worth running over the base Qwen3.8-27B? That depends on what you need. If your use case touches on politically sensitive topics where you want answers grounded in verifiable facts rather than state-aligned framing, the propaganda evaluation suggests Arete-27B is a meaningfully different model on that axis, cutting propaganda-flagged answers from 22% to under 1%. If your work is unrelated to those seven value categories, general coding, summarization, agentic tool use, the two capability benchmarks reported show no real difference between Arete-27B and the base model, so the choice is more about licensing and values alignment than raw capability. Enterprises have Remy. ⌂ Self-hosted ✓ SOC 2 ✚ HIPAA ☰ Audit log Both the model weights and the Propaganda Eval dataset are public, so you can run the same 803-prompt evaluation against Arete-27B, the base Qwen3.8-27B, or any other model to check the claims yourself rather than taking the benchmark table at face value. Frequently Asked Questions What base model is Arete-27B built on? Arete-27B is a supervised fine-tune of Alibaba’s Qwen3.8-27B, a 27 billion parameter model. Nous Research kept the architecture unchanged and trained only the language-model weights. Is Arete-27B released under Apache 2.0? The topic brief and release describe Arete-27B as an Apache 2.0 licensed open-weight model, consistent with Nous Research’s publication of both the weights and the evaluation dataset for community reuse. How much VRAM does Arete-27B need? As a dense 27B model, expect around 54GB in bf16/fp16, around 27GB at 8-bit quantization, and roughly 14 to 16GB at 4-bit quantization, plus additional memory for KV cache if you use long context the model supports up to 64k . Does Arete-27B lose performance compared to the base model? On the two benchmarks Nous Research ran, no. Hermes Bench scores were within 0.2 percentage points within noise , and Terminal-Bench-2.1 scored higher for Arete-27B 70% vs 67% , though the team cautions against reading that gain as a confirmed improvement due to possible harness differences. What exactly did the fine-tuning change? It targeted seven categories, freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy, using supervised fine-tuning on data generated by an open teacher model called Inkling. It raised factual recall and sharply cut answers that reproduced official state framing, most dramatically in the freedom of expression category.