Arete-27B Locally: 17.20 GB at Q4, 16.05 GB in MLX 4-bit

Nous Research published Arete-27B on 6 November 2026 under Apache-2.0. It is a post-train of Qwen3.8-27B that Nous says is meant to "reinforce Western values." Community quants showed up within hours: GGUF from bartowski and two MLX conversions. Because the architecture is untouched, the hardware question has a short answer. Arete-27B needs exactly what Qwen3.8-27B needs. Below are the real file sizes and the Mac or GPU tier each one fits.

What Arete-27B is

Arete-27B is a 27,356,728,560-parameter dense model, per the Hugging Face API. Its declared base is Qwen/Qwen3.8-27B and its model_type is qwen3_5. According to the model card, the training was supervised fine-tuning only: no reinforcement learning and no architectural change. The card lists about 52.8 million training tokens at 64K context over 1,202 steps.

The published config.json confirms the shape is the base model's: Qwen3_5ForConditionalGeneration, 64 layers, 4 key-value heads at a head dimension of 256, a 262,144-token maximum context, and full_attention_interval set to 4. Consequently, every memory rule that applies to Qwen3.8-27B applies here unchanged.

The checkpoint is Apache-2.0. The training pairs are public as NousResearch/arete-data-pairs, and the evaluation set as NousResearch/Propaganda_Eval.

The propaganda numbers are Nous's own claim

The announcement describes an in-house evaluation of 803 prompts across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment. According to Nous, the rate of answers it classes as propaganda falls from 22% for the base Qwen3.8-27B to 0.9% for Arete-27B. Nous also reports a fact score rising from 2.32 to 2.98 out of 5, with thinking off.

Read those as Nous's figures. The benchmark, the classifier and the comparison were all built by the team that trained the model, and no third party has reproduced them as of this check. ModelFit has not tested them either; we track hardware fit, not value alignment.

The file sizes

All sizes below come from the Hugging Face file-tree API of each repository, checked 7 November 2026. Sizes are in decimal GB.

The MLX 4-bit figure is the sum of its three safetensors shards. bartowski also ships an f16 vision projector at 0.93 GB, needed only for image input.

Which Mac runs Arete-27B

On Apple Silicon, macOS does not let the GPU wire all of the unified memory. ModelFit plans against 70% of RAM up to 32 GB, rising to 85% at 128 GB. That gives 16.8 GB usable on a 24 GB Mac, 22.4 GB on 32 GB and 34.8 GB on 48 GB.

In short, Q4_K_M or MLX 4-bit on a 32 GB Mac is the default, and 48 GB is where 8-bit becomes practical. The tiers track the Qwen3.8-27B page. One nuance: ModelFit's Qwen3.8-27B Q4 entry is 16.5 GB and just squeezes under the 24 GB budget, while Arete's Q4_K_M at 17.20 GB lands one tier up. The Arete-27B model page ranks it by machine.

Which GPU runs Arete-27B

On a discrete GPU the weights and the KV cache share the card's VRAM.

• 24 GB card (RTX 4090, RTX 3090, RX 7900 XTX): Q4_K_M at 17.20 GB leaves about 6.8 GB for context and buffers, enough for 32K tokens and more. Q5_K_M at 20.68 GB also loads, with a shorter context. Q8_0 at 28.67 GB does not fit.
• 32 GB card (RTX 5090): Q6_K at 23.62 GB is the comfortable pick. Q8_0 at 28.67 GB loads, but leaves about 3.3 GB, so keep the context short.

For the general method, the VRAM guide applies unchanged.

How much the context adds

Only 16 of the 64 layers use full attention; the other 48 are linear-attention layers with a fixed-size state. So only those 16 layers carry a KV cache that grows with length. At f16 each one stores 2 x 4 x 256 x 2 = 4,096 bytes per token, which makes 64 KiB per token across the 16 layers.

• 32,768 tokens: about 2 GiB
• 131,072 tokens: about 8 GiB
• 262,144 tokens: about 16 GiB

This is arithmetic from the published config, not a measurement, and it assumes an unquantized f16 cache. A q8_0 cache halves it.

Runtime support

The qwen3_5 architecture is supported upstream in both runtimes: llama.cpp merged it in PR #19468 on 10 February 2026, and mlx-lm in PR #869. Any recent llama.cpp, LM Studio or mlx-lm build loads Arete the same way it loads Qwen3.8-27B. There was no Ollama library entry for Arete at the time of writing, though a GGUF from bartowski can be imported by hand.

We publish no speed numbers for Arete-27B. Since the architecture and size are identical, throughput should track Qwen3.8-27B at the same quant, but that is an expectation, not a measurement.

FAQ

Is Arete-27B harder to run than Qwen3.8-27B?

No. Same architecture and the same 27B class. The Q4_K_M file is 17.20 GB against 16.5 GB for ModelFit's Qwen3.8-27B Q4 entry, so on a 24 GB Mac Arete is the tighter of the two; from 32 GB up there is no practical difference.

Can a 24 GB Mac run it?

Only at IQ4_XS (15.24 GB), Q3_K_M (13.17 GB) or MLX 4-bit (16.05 GB), and with little room for context. A 32 GB Mac is the comfortable floor for Q4_K_M.

Is there an MLX version?

Yes, two community conversions: Duckets/Arete-27B-MLX-4bit (16.05 GB) and pjain001/arete-27b-8bit-mlx (28.66 GB). Neither is from Nous.

Did Nous really cut propaganda from 22% to 0.9%?

That is what Nous reports on its own 803-prompt evaluation. It has not been independently reproduced.

Sources

• NousResearch/Arete-27B model card
• Arete-27B config.json
• Nous Research announcement
• bartowski/NousResearch_Arete-27B-GGUF
• bartowski/NousResearch_Arete-27B-GGUF file tree
• Duckets/Arete-27B-MLX-4bit
• pjain001/arete-27b-8bit-mlx
• NousResearch/Propaganda_Eval dataset
• llama.cpp PR #19468 (Qwen3.5 support)
• mlx-lm PR #869 (Qwen3.5 support)