ModelBrief · Edge0-145B · Topic Page
Last updated Nov 20, 2026

Edge0-145B Benchmarks: Mean Only +1.1 Over Base

ModelBrief · Published 2026-11-20 · Fact-check 2026-11-20 · Reference: Edge0 official blog, Hugging Face model cards, Signal Sentry

Every benchmark number for Edge0-145B today comes from a single table: six internal evaluations in the official blog. The blog says so itself — these are Edge0's internal evaluations, with the full setup, hardware environment and reproduction instructions to come with a technical report that is still unpublished as of November 20, 2026. This article reproduces that table in full, adding nothing and removing nothing, and does two things: it works out the real gap against the base model, and it says which numbers cannot be trusted yet.

All six benchmarks (vendor-reported, no independent verification)

BenchmarkEdge0-145BQwen3.8-Flash-Next (base)DeepSeek-V4-Flash-0731Claude-Opus-4.6 (Max)
DeepSWE 1.164.658.754.4—
Terminal-Bench 2.186.5283.382.778.2
HLE35.435.933.840.0
GPQA-Diamond90.491.790.891.3
IFBench80.681.379.262.5
AA-LCR79.379.179.778.0
Mean (6)72.871.770.1—

All six scores are vendor-reported (Edge0 internal evaluations) with no independent verification as of Nov 20, 2026. The blog table's Flint, Muse Glimmer and Gemma 4 rows are taken from Recursive Flint's evaluation (Terminal-Bench / IFBench) and Artificial Analysis (HLE / GPQA / AA-LCR), not from Edge0's own testing — do not read them as Edge0's scores.

An honest reading: the lead is real, and small

The coding and agent leads are real. On DeepSWE 1.1, Edge0-145B beats its base by 5.9 points (64.6 vs 58.7); on Terminal-Bench 2.1 by 3.2 points (86.52 vs 83.3); the lead over DeepSeek-V4-Flash is larger still (10.2 points on DeepSWE, 3.82 on Terminal-Bench). These two are exactly Edge0-145B's positioning — local coding and agent work — where the MoE rebuild and post-training paid off.

But the mean lead over the base is only +1.1 points. 72.8 vs 71.7 — across six benchmarks, the rebuild and post-training bought 1.1 points. And the other side matters more: three scores sit below the base — HLE 35.4 vs 35.9, GPQA-Diamond 90.4 vs 91.7, IFBench 80.6 vs 81.3. All three knowledge and instruction-following scores trail, which means this rebuild did not lift every capability — it tilted capability toward coding and agents. AA-LCR is essentially level (79.3 vs 79.1).

Discount the Claude-Opus-4.6 column. The blog table sets Edge0-145B's Terminal-Bench 2.1 (86.52) against Claude-Opus-4.6 (78.2); commenters have already pointed out that this comparison is outdated — the Opus 4.6 figure is not a same-period, same-protocol measurement. Before citing "beats Claude-Opus-4.6", say that the column has a timeliness problem.

Independent verification: zero as of Nov 20

Signal Sentry (Nov 13–15) puts it plainly: "No independent results yet. They appear here as Epoch AI and LMArena add this model." Our check agrees: Artificial Analysis has no Edge0-145B entry — the AA numbers borrowed in the blog table belong to other models (Flint / Glimmer / Gemma 4), not a score AA gave Edge0-145B; LiveBench has no run; LMArena has no listing. A YouTube deep-dive from Nov 15 comments on the vendor numbers; it is not an independent run, so do not cite it as verification.

The good news is that verification can start now: the weights and code are public, so anyone can run them. The two things on the watchlist are the technical report and the first independent runs (an AA, LiveBench or LMArena listing) — either would materially change how much this table can be trusted.

Speed figures: also vendor-measured, also conditional

Build (RTX 5090 D 24GB, vendor-measured)Agent decode16K prefill
NVFP498.3 tok/s3,165.7 tok/s
Q3_K121.2 tok/s3,554.9 tok/s

The blog's headline figures are 80+ tok/s decode, 1,100+ tok/s prefill, ~120W typical power. Like the benchmark scores, the speed figures are Edge0's own measurements with no independent reproduction yet; quote them with the build name and the hardware (RTX 5090 D 24GB), not as a single number.

ModelBrief's take: This table deserves a careful read, not blind trust. The coding and agent leads are a real signal — +5.9 on DeepSWE and +3.2 on Terminal-Bench are not small; but a +1.1 mean and three scores below the base show the rebuild did not work everywhere. The bigger problem is provenance: all six are vendor-reported, the technical report is still unpublished, and AA, LiveBench and LMArena have none of them listed. Until independent runs arrive, any "frontier-level" conclusion needs the words "vendor-reported" attached.