So the question is never "is it clever?" but "can a step reach it, and may we run it?" Both answers come from licence and hosting. **49% of the top open-weight models are permissive**, 24% are open with conditions, and 27% are API only, in the [API Price Index](https://apipriceindex.com/guides/open-weight-licenses/) breakdown, meaning licensing constraints will affect the majority of available models.  Roughly half can be self-hosted without legal review. Another quarter needs someone to read the conditions first, and over a quarter you can only reach through the maker's endpoint, so your workflow depends on their uptime and pricing. ### What the model page does and does not say Across a wider set, commercial-use terms are murkier. [LocalLLM](https://localllm.az-arte.com/en/licenses) counts 826 models cleared for commercial use, 239 conditional, 172 unverified and 28 non-commercial.  **The 172 unverified are the important figure**: for them, nobody has confirmed the terms, so a team that wires one in is guessing about its own legal position. For Arete 27B, five items on Nous Research's page decide whether it can be used: licence terms, weight access, who hosts it, context length, and whether an API endpoint exists. The model page confirms the 64k training context and benchmark results, so we mark those as stated; the licence is not stated there. **Treat it as you would any model in that 172, and don't build a production step around it.** ## What Arete 27B's maker claims According to [Nous Research](https://huggingface.co/NousResearch/Arete-27B), the new thing is a post-training pass: it took Alibaba's Qwen3.8-27B, a model trained in China, and tuned it "to reinforce western values." That's the whole claim, and it concerns behaviour, not new architecture.  ### Facts stated on the model page "We post-trained over Alibaba's Qwen3.8-27B **to reinforce western values on a Chinese-trained model**. We call this new model Arete-27B." That is the page's own wording. The 27B in the name matches the base, so this is the same size class as the model it was built from, not a larger or smaller one. For a workflow builder, that means any hosting or cost reasoning you already do for a 27-billion-parameter model is the starting point, though you would still have to check it yourself. Compare what the page gives with what a builder would want to see, in the table below. Five of six rows have entries, all from the model page. | Fact | As stated by Nous Research | Link or not stated | |---|---|---| | Base model | Alibaba's Qwen3.8-27B | [model page](https://huggingface.co/NousResearch/Arete-27B) | | Size | 27B, in the model name | [model page](https://huggingface.co/NousResearch/Arete-27B) | | Context length | 64k training context | [model page](https://huggingface.co/NousResearch/Arete-27B) | | Fact score | Rose from 2.32 to 2.98 out of 5 | [model page](https://huggingface.co/NousResearch/Arete-27B) | | Propaganda answers | Cut from 22% to 0.9% | [model page](https://huggingface.co/NousResearch/Arete-27B) | | Licence | Not stated | not stated | _Prices and plan limits checked against [huggingface.co](https://huggingface.co/NousResearch/Arete-27B) on November 10, 2026._ Only the licence row is empty, and it is one of the rows that decide whether a model can sit in a workflow step. ### What Arete 27B's listing leaves out  Look at the model page and you find a 64k training context and benchmark results, but no licence terms. Without context length you can't say how much of a ticket thread or document batch fits in one call. Without a licence you can't say whether a client deployment is allowed. The page's own benchmarks are the maker's claims: fact score 2.32 to 2.98 out of 5, and Hermes Bench 65.77% versus 65.60%. Until the maker supplies those fields, treat each one as an open question, and don't budget a workflow around a number you haven't seen. ## What integration paths does Arete 27B offer? Wire it in through general building blocks. Host the model yourself, put an OpenAI-compatible endpoint in front of it, then call that endpoint from an HTTP step, an AI step, or an MCP server. Try them in this order, because each adds moving parts: 1. Host the model on a machine you control, since nobody offers it as a ready-made service you can assume. 2. Expose an OpenAI-compatible endpoint from that host, so anything that speaks the common chat format can reach it. 3. Call the endpoint from a plain HTTP request step and confirm a round trip works. 4. Move to an AI step with a custom provider pointed at the same endpoint. 5. Add an MCP server only if the model needs to call tools. It's a stand-in, not Arete 27B's own requirement, so read it as the shape of the cost: lower precision shrinks the memory bill, and that decides whether the model fits on one GPU or needs several. Whatever precision you settle on fixes the hardware, and the three options that follow differ only in how the workflow reaches it. ### Option 1: an HTTP request step Start here. You send a POST to the chat completions path on your endpoint, with the model name and a messages array in the body, and map the reply's first choice into the next step. You see the raw request and response, so when the output is wrong you know whether the fault is the model or your field mapping. The cost is that you hand-build every call and parse every answer yourself. ### Using an OpenAI-compatible AI step for Arete 27B If your platform's AI step lets you set a custom base URL, API key and model name, point it at your endpoint. You get the platform's native prompt and output handling instead of hand-built JSON. Check before committing that the step allows a custom base URL at all; some AI steps are locked to a fixed list of hosted vendors, and then Option 1 is your answer. Reach for an MCP server only when the model must act: query a database, create a ticket, read a file. An MCP server exposes those actions as named tools the model can call. It does nothing for plain text generation, so skip it until Options 1 and 2 work. ## What Arete 27B's cost and speed mean Nobody has published a cost or a speed for Arete 27B. The only honest numbers come from its closest stand-in, Qwen3.8-27B, and the table below shows how little of the Arete row can be filled in. | | Cost | Speed | Source | |---|---|---|---| | Hosted API model (Qwen3.8-27B) | $0.05-$0.99 per 1M input tokens | Not stated by the provider | Vendor pricing pages | | Self-hosted open-weight model like Arete 27B | Not stated | Not stated | None published | | Current model | Set per vendor | Set per vendor | Vendor model pages | Read the empty cells as the finding: you can price a hosted 27B call today, and you can't price Arete. **$0.05 to $0.99 per 1M tokens** is the range for hosted Qwen3.8-27B input, a twenty-fold spread, so the provider you pick matters more than the model size. Nous Research has stated no Arete price, and self-hosting adds a GPU bill the model card won't show. Any budget line for Arete is a guess until someone runs it. ### What we know about Arete 27B speed We haven't run Arete, and no latency figure for it exists. A 27B model of similar size gives a rough ceiling.  [GPUBattle](https://gpubattle.com/guides/qwen3-8-27b-gpu-benchmarks) measured Qwen3.8-27B decoding at 80.5 tokens per second on an H100 80GB, which means a 400-token reply takes about five seconds on data-centre hardware. On consumer cards the figures sit close behind or well below that. The RTX 5090 managed 75.74 tokens per second, so a card far cheaper than an H100 gives up only about six percent. The RTX 3090 managed 41.93, roughly half the H100's rate, which turns the same 400-token reply into nearly ten seconds. That's tolerable for a nightly batch step and painful for a workflow where a person waits on the answer.  Because Arete may quantize, batch or behave differently, these figures are a stand-in. Treat them as a reason to run your own test on your own hardware before committing a step to it. ## What Activepieces does about this Activepieces has shipped nothing specific to Arete 27B, and no one here has tested the model, so this section only covers what the platform lets you do with it today. The Webhook and HTTP integrations are the two the article leans on, and both sit near the top of the library when you sort by popularity. Once you have an OpenAI-compatible endpoint in front of your own host, the HTTP integration can send the chat request and return the response to the next step. That is Option 1 above, and you can run it before anyone writes an integration. The same reasoning applies to cost. On the Plus plan at $20 a month you bring your own AI keys. The tokens, or the GPU time if you self-host, are billed to you at your own rate, which is the number the model page leaves empty. Activepieces can't fill that cell in for you, but it doesn't add a markup to it either. The platform also doesn't decide which model you run. A tool that resells its own model has made that choice before you log in. Here the choice is yours, so you can run Arete 27B beside a model you already trust and compare the outputs on your own data. That comparison covers the properties the model card doesn't state: clean JSON on the fifth retry, how long a reply takes on your hardware, and whether the licence lets you ship it to a client. MoneyGram, Moneypenny, Alan and FundingSocieties run Activepieces, though none of them has told us they use Arete. If you want to act on this, start with one HTTP step against a test endpoint, and keep the step out of production flows until Nous Research states the licence and context length. If the model needs to call tools later, you can reach Activepieces from Claude, ChatGPT, Cursor or any MCP client, which covers the last step of the ladder without building anything new. ## Related reading - [Harness2: Google Research Repository Overview (2026)](https://www.activepieces.com/blog/harness2-google-research-repository-overview-2026) - [10 Best NotebookLM Alternatives for AI Research (2026)](https://www.activepieces.com/blog/10-best-notebooklm-alternatives-for-ai-research-2026) ## References - [API Price Index](https://apipriceindex.com/guides/open-weight-licenses/) - [GPUBattle](https://gpubattle.com/guides/qwen3-8-27b-gpu-benchmarks) - [LocalLLM](https://localllm.az-arte.com/en/licenses)Until a workflow step can call it, a model doesn't matter to an automation team, and that depends on the terms and access the maker publishes, not on benchmarks.