ModelBrief · Nex-N2.6 · Topic Page
Last updated Nov 20, 2026

Nex-N2.6 Benchmarks, All of Them: OSWorld 83.0 Leads — and the Two Tests It Loses

ModelBrief · Published 2026-11-16 · Fact-check 2026-11-16 · Reference: Hugging Face model card

Nex-N2.6's benchmark data comes in one batch only: three benchmarks and one internal reliability table on the vendor model card. No Artificial Analysis, no LiveBench, no LMArena — every number comes from Nex-AGI itself. This article takes the three benchmarks apart, shows which one is a genuine win and which two lose to the base and a rival, then lists the only five pieces of independent evidence.

The three-benchmark table: OSWorld is the only clear win

The model card's table (all vendor-reported; rival scores as Nex-AGI lists them, not each vendor's published values): OSWorld 83.0 (predecessor 82.2, base GLM-5.3-Flash 78.0, DeepSeek-V4.1-Flash 71.7), first in the table; Terminal-Bench 2.1 84.1 (predecessor 82.7, base 84.3, DeepSeek 82.7), 0.2 below its own base; AutomationBench 45.7 (predecessor 44.2, base 48.8, DeepSeek 54.8), 3.1 below the base and 9.1 below DeepSeek. Only OSWorld is a clear win — and only 0.8 above the predecessor. The headline number's real weight has to be read in that margin.

Nex-N2.6 model-card benchmarks (vendor-reported) Nex-N2.6 Nex-N2.5-Pro GLM-5.3-Flash DeepSeek V4.1 OSWorld 83.0 82.2 78.0 71.7 Terminal-Bench 2.1 84.1 82.7 84.3 82.7 AutomationBench 45.7 44.2 48.8 54.8
Model-card benchmarks (points; Nex-AGI model card, Nov 10, 2026; all vendor-reported; orange is Nex-N2.6, teal is the GLM-5.3-Flash base, grey is N2.5-Pro and DeepSeek-V4.1-Flash)

The reliability table: repetition 16.0% to 1.6% is a real gain

The card also carries an internal reliability evaluation (the vendor's own known-issue replay): reasoning repetition N2.5-Pro 16.0%, GLM-5.3-Flash 6.0%, N2.6 1.6%; overall non-repetition 91.0% / 96.2% / 97.8%; valid tool selection (overall) 82.2% / 85.3% / 100.0%; (when with tools) 80.5% / 84.1% / 100.0%. This is post-training's strongest result: repetition largely fixed, tool selection perfect on replay. But mind the basis: this is the vendor replaying its own known issues, not a third-party test — and lininn.cn's failure case (next section) happened exactly in the kind of real long task a replay doesn't cover.

The harness trap: model-level scores, not FP8-export scores

The card's own words: "The following results are reported by the Nex-AGI team for this release. Scores depend on the evaluation harness and settings. These are model-level results, not a separate evaluation of this FP8 export." Each sentence matters: the scores are vendor-reported; they vary with harness and settings; the downloadable FP8 weights have never been independently tested. Any claim that 83.0 is "the score the open weights achieved" is wrong. Equally unwritable: a single throughput/latency figure — Hokai claims 80 tok/s at 0.89s, OrcaRouter ~35 tok/s at p50 2.81s, more than 2× apart; this site adopts neither.

Independent evidence: five records, no score

As of Nov 16, Artificial Analysis has no N2.6 model page, LiveBench has no run, LMArena has no listing; Epoch AI says there are no standardized scores yet. The only independent evidence is five records: Plotly DataAnalyticsBench 38/43 (lininn.cn, Nov 12), tied with the base GLM-5.3-Flash — post-training opened no gap on this board; on the same board as this site tracks it, Carbon scores 42/43 and Hydrogen 43/43 (full marks), ahead of N2.6. SMF Cold Iron: 76.4% (120/157) (SMF Clearinghouse, Nov 12), No. 24 on its board — up 24.8 points on predecessor N2.5-Pro's 51.6%; but coding is only 11/30, and 14 of the 17 failures are SyntaxErrors (including em-dashes written into code); a single run on a 157-test suite, and the model cannot turn thinking off, so this ran at medium effort. GameBench 2: 72.7 (Low), No. 22 of 95 (llmbenchmark.ai, Nov 15), one board, a low-effort run. hypeorcode: 6 runs, 28m48s, $0.077 — too small a sample to conclude anything. The lininn.cn failure case (Nov 12): in a third-party test, the model ran 1,670 calls over 2h20m on one data-analysis question without finding the answer; the bill was $10.19 — the strongest critical data point available.

The weak-spot list: the gaps beyond the three benchmarks

The weak spots, collected: Terminal-Bench 2.1 below its own base; AutomationBench below the base and DeepSeek; an OSWorld lead over the predecessor of only 0.8; no independent score of any kind; the FP8 export untested; the Plotly independent benchmark run tied with the base; a real long task that ran 1,670 calls without an answer. These don't invalidate 83.0 — but they set N2.6's boundaries: computer-use tasks are worth trying, terminal and business-process tasks may not be its best use, and long tasks need a call cap.

FAQ

What does OSWorld 83.0 mean?

First in the model card's table, 5.0 above the base, 0.8 above the predecessor, 11.3 above DeepSeek-V4.1-Flash. But all vendor-reported, with no independent reproduction; and the lead over the predecessor is only 0.8 points — read the headline in that margin.

Why say the FP8 weights were never tested?

The card's own words: the scores are model-level results, not a separate evaluation of the FP8 export. The downloadable FP8 weights have had no independent benchmark run, so 83.0 cannot be read as the export's score.

Is there any independent score?

No. As of Nov 16, Artificial Analysis, LiveBench and LMArena have all yet to list N2.6; Epoch AI says there are no standardized scores yet. The only independent evidence is Plotly 38/43, SMF Cold Iron 76.4%, GameBench 2 72.7, hypeorcode's 6 runs, and lininn.cn's failure case.

ModelBrief's take: N2.6's benchmark profile is remarkably consistent: computer use and reliability are the strengths, terminal and business-process work the weaknesses, and evidence depth the hard limit. The official and independent numbers don't disagree on direction — they disagree on existence: 83.0 is what the vendor says, while the independent world has only 38/43, 76.4%, and one 1,670-call failure. The key to reading N2.6's scores isn't memorizing 83.0 — it's remembering who measured it, and that nobody has measured the FP8 weights.