On Tuesday, November 10, 2026, a new model quietly appeared on OpenRouter: Nex-N2.6 — no launch post, no model card, no benchmark table, and the vendor's own website was still promoting the previous generation. Two days later, the weights and model card became readable. This was not one release but two: the Nov 10 API listing, and the open-sourcing around Nov 12. This article consolidates everything confirmed at launch: the timeline, the base swap, the Nex-AGI background, and why the earliest reviews went stale within a day.
On Nov 10, OpenRouter listed nex-agi/nex-n2.6: callable, but with no documentation at all — no launch post, no readable model card, no benchmark table. The GitHub repo sii-nex/Nex-N2.6 was created the same day. On Nov 11, nothing had improved: the Hugging Face repo nex-agi/Nex-N2.6 was still not publicly readable — OrcaRouter hit a login prompt, Hokai a missing page — and ai345 (Chinese) found ModelScope 404 and HF unreachable. Around Nov 12, the weights and model card finally became publicly readable: Open Model Weights published its weights report that day, and lininn.cn published its critical long-form piece. On Nov 14, AI Weekly published its release summary, and the After Intelligence newsletter listed N2.6 as one of four lead frontier items. "Released Nov 10" is accurate only as "API listed Nov 10, open-sourced ~Nov 12" — don't merge the two dates.
The predecessor Nex-N2.5-Pro (released Sep 8, 2026) used a Qwen3.5 base at ~397B parameters; N2.6 switched to Z.ai's GLM-5.3-Flash (released Aug 26, 2026, 320B/18B). That is not a small change: a base swap replaces the tokenizer, the architecture details and the capability distribution wholesale. The parameter count changed with it: 321.3B total, ~18B active, a 45-layer, 4,096-dim backbone (3 dense FFN + 42 sparse MoE layers). The 1B above the base's 320B is a rounding/adapter artefact, not a new architecture. The cost of the swap is written in the price: N2.6's $0.12/$0.40 is 60% above N2.5-Pro's $0.075/$0.25 — a base swap and a price hike together; see our pricing guide.
Nex-AGI is a Shanghai team, and N2.6 was open-sourced by the Shanghai Innovation Institute under the GitHub org sii-nex. Community channels are Feishu and WeChat (QR codes on the model card). The launch is small in scale: 14 GitHub stars (as of ~Nov 12), 410 Hugging Face downloads in the past month (AI Weekly, Nov 14). As of Nov 16, no tier-1 Chinese tech outlet (机器之心, 新智元, 量子位, IT之家) has covered it, and no mainstream English outlet (TechCrunch, The Verge, Reuters, Bloomberg, Wired) either — the publisher is a Shanghai lab, but Chinese coverage is just two pieces, ai345 and lininn.cn.
OrcaRouter's and Hokai's Nov 11 analyses were the first English reviews after launch — and the fastest to go stale. Written before the model card was readable, both said "no benchmarks, no weights"; when the weights became readable on Nov 12, both claims became wrong. Cite them selectively: their price and base-swap analysis still stands, but the "no benchmarks / no weights" conclusion should not be used. Their throughput figures also conflict (Hokai 80 tok/s, OrcaRouter ~35 tok/s); this site adopts no single number. The lesson is simple: for a silently launched model, the first reviews describe what was invisible at the time, not what doesn't exist.
Because Nov 10 was only the silent OpenRouter listing, with no model card or weights published alongside; the HF repo was still unreadable on Nov 11 and became readable only around Nov 12. The release date and the open-weights date are two different things, and this site writes them separately.
A Shanghai team; N2.6 was open-sourced by the Shanghai Innovation Institute (GitHub org sii-nex), with community channels on Feishu and WeChat. Small in scale: 14 GitHub stars, 410 HF downloads a month; as of Nov 16, no mainstream media in either language has covered it.
Reference: Hugging Face — nex-agi/Nex-N2.6 model card · OpenRouter — nex-agi/nex-n2.6 model page (rewritten and organized by ModelBrief; facts per the model card and OpenRouter)