On 2026-11-06 Nous Research released Arete-27B, an ~52.8M-token SFT of Alibaba Qwen3.8-27B aimed at Western-aligned answers. On Nous's 803-prompt seven-category eval, vendor figures show propaganda answers ~22% to ~1% and fact mean ~2.32 to ~2.98/5; Hermes Bench ~65.60% vs base ~65.77%. Weights, eval, and data are public.
Key Takeaways
- ✓Source: Nous blog Introducing Arete-27B (Nov 2026) + @NousResearch post
- ✓Method: plain SFT on Qwen3.8-27B (~52.8M tokens, 2 epochs, no RL/PO); teacher Inkling; vision encoder frozen
- ✓Eval (vendor): 803 prompts / 7 value categories; propaganda rate ~22% to ~1%; fact mean ~2.32 to ~2.98/5; judge Grok 4.7
- ✓Capability: Hermes Bench ~65.60% vs base ~65.77%; Terminal-Bench-2.1 vendor ~70% vs base 67%
- ✓Open: Apache 2.0 weights, Propaganda_Eval, and training pairs on Hugging Face
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Core Background & Industry Pain Points
Strong Chinese open bases often ship with state-framed answers, raising adoption friction for federal and enterprise use. On 2026-11-06 Nous Research released Arete-27B, arguing a shallow SFT can cut propaganda-style framing toward Western open-model levels while keeping agentic skill.
Architecture Highlights & Internals
Base: Alibaba Qwen3.8-27B. Training: ~52.8M tokens, batch 48, 1,202 steps over 2 epochs; LM weights trained, vision encoder frozen, MTP disabled. Teacher: Inkling. Data spans seven value categories after dedupe and eval-overlap removal.
Authoritative Benchmarks & Measured Scores
Vendor 803-prompt eval: propaganda rate ~22% to ~1%; fact mean ~2.32 to ~2.98/5; judge Grok 4.7; no independent replication yet. Hermes Bench ~65.60% vs base ~65.77%; Terminal-Bench-2.1 vendor ~70% vs 67%. Treat vendor eval as self-reported until outsiders rerun it.
Developer Hands-on Guide
Weights: Hugging Face NousResearch/Arete-27B (Apache 2.0); eval and data pairs public. Useful for local/self-hosted Qwen derivatives needing Western-aligned framing; rerun Propaganda_Eval before production and do not treat the vendor propaganda rate as a general safety proof.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.