← All model news

Nous releases Arete-27B: pure SFT on Qwen3.8-27B cuts self-scored promotional answers from 22% to 0.9%

According to a MarkTechPost report dated November 7, 2026, Nous Research released Arete-27B, which adapts Alibaba's Qwen3.8-27B with supervised fine-tuning alone. In its 803-prompt self-evaluation, promotional answers fell from 22% to 0.9% and the factual mean rose from 2.32 to 2.98. The weights are released under Apache 2.0, but no hosted inference provider currently offers the model, according to the report.

Published Updated

What changed

According to a MarkTechPost report dated November 7, 2026 (by Asif Razzaq), Nous Research released Arete-27B, a 27-billion-parameter dense open-weight model produced by supervised fine-tuning Alibaba's Qwen3.8-27B. The report says the adaptation used SFT only, with no reinforcement learning, no preference optimization and no architectural changes. The weights are released under Apache 2.0, the same as the base model. Training covered roughly 52.8 million tokens, a 64k context, 2 epochs, 1202 steps and a batch size of 48, with a peak learning rate of 1e-5. All language-model weights were trained, while the vision encoder was frozen because the data was text-only and the multi-token prediction head was disabled. The teacher model was Inkling, an open model from Thinking Machines. Nous said it chose an open teacher to avoid distillation usage-term restrictions, and did not use Nemotron 3 Ultra because of that model's record on China-related topics. The accompanying dataset, arete-data-pairs, contains 34,878 prompts across 7 topics, with both Inkling and Qwen answers for each prompt. Of these, 29,236 accepted teacher answers make up the SFT training set. The evaluation set, Propaganda_Eval, contains 803 prompts spanning 7 value categories and was scored by Grok 4.7. It reports a factual mean from 0 to 5 and the share of answers containing "any propaganda." The scoring rules require quoting exact wording before marking an answer as propaganda, training prompts overlapping with the evaluation set were removed, and the evaluation set is released under MIT with a scoring script. The reported results: the factual mean rose from 2.32 to 2.98, the propaganda share fell from 22% to 0.9%, and answers scoring 0 fell from 359 of 803 to 216. The largest change was in the free-speech category, where the factual score rose from 0.89 to 2.44 and propaganda fell from 60% to 1%.

What it means for developers

What developers should really note here is the method rather than any single score: SFT alone, without architectural changes, can substantially shift a class of answer behavior in the vendor's self-evaluation, and the report says capabilities were largely preserved. Hermes Bench went from 65.77% to 65.60%, which Nous itself called noise, and Terminal-Bench 2.1 rose from 67% to 70% in Nous's own test harness, a gain Nous did not attribute to its training. In other words, changing value orientation looks cheaper than expected, but the costs show up elsewhere: the factual mean of 2.98 is still below the 3.02 shown for Muse Glimmer 30B in the report's table, and Arete's 0.9% propaganda share is above the 0.5% for Muse Glimmer and Gemma 4 31B, while 216 of the 803 answers still scored 0. That means "repeating official talking points less" and "getting facts right" are two separate dimensions, and one metric cannot substitute for the other. On deployment, the report says no hosted inference provider currently offers the model. The BF16 weights are 53.81 GB, a community Q4_K_M quantization is 17.20 GB, and it can run in llama.cpp, LM Studio and Ollama; the GGUF version also includes an image projector. But the base Qwen3.8-27B natively supports 262,144 tokens, while the fine-tuned version's context is listed as "not disclosed (trained at 64k)," and there is little data on how the long-context and image paths actually perform. On licensing, the report says the weights are Apache 2.0 and the evaluation set is MIT; the specific terms and commercial boundaries should still be checked against the model card and dataset page before use. Also, much of the hidden cost of swapping models lies in prompts and system messages: if you relied on the base model's refusal behavior as a safety backstop, refusal rates will change with this fine-tuned version, and the backstop logic needs to be adjusted accordingly.

Integration and evaluation

The following are conditional suggestions and inferences, not reported facts. First, whether to replace the local base model with Arete-27B. The reported change is a large drop in a class of behavior in self-evaluation, but the capability figures come from different setups: under Qwen's official methodology the base model scores 73.0 on Terminal-Bench 2.1, while the 67% and 70% come from Nous's test harness, so they cannot be subtracted directly. Only if your workload is terminal or coding agent tasks, you must self-host, and you can accept having no hosted service is it worth putting Arete-27B on your candidate list. If you depend on a hosted API or need official benchmark backing, stay with the base model for now: the base Qwen3.8 27B is already listed in the site catalog, and pricing and quotas are subject to the site's pricing page. As a decision criterion, run both models on a small set of your own real requests (for example, 50 to 200 historical tickets or scripted tasks), compare task pass rates and refusal rates, and set an acceptable threshold for pass-rate decline before deciding. Second, whether to adopt Nous's evaluation method as a release gate. The usable change is that the evaluation set is released under MIT with a scoring script, but scoring depends on a third-party closed model like Grok 4.7, and results may not be consistent with a different judge. The precondition is that you have a judge model with comparable instruction-following ability and that your live traffic actually falls within the 7 categories it covers. The approach: first run a consistency check, manually label 50 to 100 samples and compare them one by one with the judge's results, and only use it for batch regression once agreement reaches a threshold you set (for example, 90%); otherwise treat it only as directional guidance. Third, choosing quantization levels and modality paths. The reported change is that a 24GB consumer GPU can run the community Q4_K_M version at 17.20 GB, and that the vision encoder was frozen on text-only data. If your use case needs image input or more than 64k of effective context, first run the quantized and BF16 versions on the same batch of prompts, measuring behavior drift after quantization and where long text gets truncated, then decide whether to lower the quantization level or drop the model.

Limits and open questions

These numbers come from an evaluation run by Nous Research itself, with Grok 4.7 as the judge. The report mentions no third-party reproduction or blind test, so the 22% to 0.9% comparison should be treated as a vendor self-evaluation, not an independently verified conclusion. Cross-model comparisons in the table are also not fully controlled: Arete and Qwen ran at temperature 0, while Muse Glimmer and Gemma ran at 0.6, and the Terminal-Bench figures come from different test harnesses. The fine-tuned version's context length is not disclosed, there is little data on long-context and image-path behavior, and the factual mean of 2.98 still trails comparable models. As of this catalog check, Arete-27B does not appear in KimiSeek's model catalog; the weight download and local-run methods described in the report point to Nous Research's release channels and community quantizations, and availability should be confirmed with the vendor or any later providers. The base Qwen3.8 27B in the catalog is a different version, and its listing does not mean Arete-27B is available. The report also says no hosted inference provider currently offers the model, so any API access availability likewise refers to the vendor side. In addition, adjustments to this kind of content orientation involve compliance and content-policy requirements in different markets, which need independent assessment before deployment; this article is not compliance advice.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

Nous Research 发布 Arete-27B:把 Qwen3.8-27B 的宣传性回答从 22% 降到 0.9%

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.