MEASUREDNEWWindows edition 1.2.1 · RTX 3090 · 24 GB · Oct 18–19, 2026
Windows PC, RTX 3090
The Windows edition on a gaming PC in daily use: the Mac campaign’s harnesses, ported and run against the installed app. Twelve catalog fleets best-of-4, the 156-job modes grid, bulk, watchers, a crash mid-chain, a three-hour sustained run, heat, energy and the Mac’s release gates.
The twelve catalog fleets on the Windows PC: winners correct of 9 jobs, median seconds per job, and the same fleet’s median seconds per job on the 16 GB Mac| Fleet | Winners, correct of 9 | PC s, median seconds per job | Mac s, median seconds per job on the 16 GB M5 |
|---|
| starter-16 | 8/9 | 3.7 | 12 |
|---|
| percept-lite | 9/9 | 9.1 | 6 |
|---|
| asr-factory | 6/9 | 10.3 | 30 |
|---|
| xfer-lite | 8/9 | 6.1 | 9 |
|---|
| ext-lite | 7/9 | 3.3 | 379 |
|---|
| agentcode-lite | 9/9 | 9.4 | 400 |
|---|
| migrate-lite | 9/9 | 6.2 | 141 |
|---|
| testgen-lite | 9/9 | 33.2 | 444 |
|---|
| mdcode-lite | 9/9 | 6.2 | 352 |
|---|
| ops-lite | 8/9 | 27.6 | 472 |
|---|
| web-bulk-lite | 9/9 | 9.2 | 401 |
|---|
| ext-core | 9/9 | 21.2 | 84 |
|---|
- 3-hour run
- 553/553
- Release gates
- 7/7
- GPU Wh per job
- 0.3–0.4
Every job completed and the judge panel ran on all 108; the crowned winner was correct 93% of the time, against 87% for a random attempt. Eight fleets kept every lane on the card; the four that rotated spent about 21 s of each job switching models. Three hours of best-of-6 jobs ran at 185, 188 and 180 an hour with lane memory flat at 20.3 GB. Decode: 190–355 tok/s on a 1B, 50–70 on an 8B.
One PC, one card; a 12 GB or 8 GB card would rotate most fleets. Shown a correct candidate, the 3B judge crowned another in 19 of 42 grid jobs. 10 of 12 generated apps were “verified” and 6 worked. The 8B curator rarely got room beside the fleet: 133 of 553 jobs got a synthesis. No image, video or music weights were installed, and the 7 fleets that need 12–14B models were not run.
MEASUREDNEWM1–M4 edition on an Apple M5 · 16 GB · Oct 14–15, 2026
Mac M1–M4 edition
The Mac app for M1 to M4 chips runs eight small models (1–4B) on the GPU through WebGPU. Each was timed and graded on 34 prompts inside the shipped app, alongside panels, the memory budget, the local OpenAI endpoint, engine jobs and the cloud relay. Llama 3.2 3B scored best (32 of 34); Llama 3.2 1B was fastest (90 tok/s); Qwen 2.5 Coder 3B passed all six code tests.
- Engine restarts
- 19 → 0
- Problems fixed
- 7/7
- Cold boot
- ~290 ms
WebKit was killing the engine whenever the window was covered or closed, taking every running job with it; the window now parks instead. A request naming an on-device model could send one step to a cloud model; those requests are now local-only. A WebLLM patch cut each load’s time over WebKit’s memory line from about 0.85 s to 0.05 s.
Measured on an M5, newer than the Macs this edition targets; the 8 GB and 24 GB tiers were simulated, and at 8 GB Phi 3.5 is refused before download. On 1–1.5B models, deliberate jobs hit the 24-step limit. A development build: ad-hoc signed, not notarized.
MEASUREDNEWDGX Spark GB10 · Oct 10, 2026
3D bench
Three open image-to-3D models on 16 test stills, then 14 Nightshift 3D desks run as real nights. Every lift is finished and painted by Meterless and checked twice.
Image-to-3D models on the bench: median lift time, peak memory and mean grade out of 5| Model | Lift, median | Memory, peak | Grade, mean of 5 |
|---|
| TripoSR | 2.0 s | 4.7 GB | 3.3 |
|---|
| TripoSG 1.5B | 167 s | 11.6 GB | 4.1 |
|---|
| Hunyuan3D 2.1 | 85 s | 12.3 GB | 4.3 |
|---|
- Valid GLBs
- 246/246
- Median finish
- 8.3 s
- Paint + finish
- 104 s
Hunyuan3D shut the box down 4 times unpaced in a warm room; with cool-down pacing it ran 9 jobs in a row at 83 °C or less. Its Tencent community licence excludes the EU, UK and South Korea. Closed and manifold: 82 of 127 raw lifts.
INTERIMNEWClaude Opus 5.5 + Haiku 4.5 · Oct 12, 2026
Agent frameworks
Not Nightshift: the Meterless engines and the most-shared agent frameworks, each against a plain same-model baseline on 11 tasks with hidden graders.
- Meterless context layer · Haiku
- +4.29
- Meterless prompt cards · Haiku
- +3.36
- Meterless app · Opus cost
- −10%
On Opus the baseline scores 99.06 and no framework moves quality clearly; there the differences are cost and time (ponytail ran 35% cheaper). The Meterless app held Opus quality (99.9) at $0.65 a run.
The same context layer on Opus scored −2.79 (not significant) at +8% cost, and the Meterless cards with an audit step scored −1.38 on Haiku at +486% cost. 675 of 774 runs graded.
MEASUREDNEWDGX Spark · Oct 1–3, 2026
Job modes
What each Nightshift mode is good for, and what it was not doing: selectors, judges, deliberate, batch, bulk, deadlines, app generation and repair.
- Deliberate, correct
- 0 → 8/15
- Gate repair
- 6/6
- Refine, seeded bugs
- 6/6
Bulk was the most dependable mode: 220 of 220 units correct with a lossless pause. Model-free conversion ran at 6,700–8,300 files an hour. 11 of 11 malformed requests were refused at submit, with a reason.
12 of 12 generated apps passed “verified”, but only 6 of 12 worked. Auto planned 256 attempts for answers a best-of-4 found in under 2 minutes. Electricity was not measured.
MEASUREDNEWApple M5 · 16 GB · Sep 22–26, 2026
Mac M5, 16 GB
Every installed catalog fleet, best-of-4, on a fanless laptop with its owner’s apps open: code, fact and extraction jobs, plus media.
- Code winners
- 21/21
- Whisper, real time
- ~32×
- SD-Turbo 512²
- 3.0 s
Plans follow the memory actually free: three rotated jobs fell from 1,201 s to 273 s. Model switches took a median 1.8–2.5 s.
14B models need about 11.4 GB free and were never admitted at 8–9 GB. Image edit and video do not fit in 16 GB. The Mac build was ad-hoc signed, not notarized.
MEASUREDOne desk-side machine · Sep 26–28, 2026
Knowledge work
The morning desk: real business briefs raced overnight by open models, read by a judge, re-checked by proofs and confirmed against an answer key.
- Deliverables
- 18
- Words, checked
- 15,424
- Per run, by build
- 0 → 11
Qwen3.8 27B won 12 of the 18; Qwen2.5 14B won 6. The security RFP answers improved build over build: judge 0.85 → 0.92 → 0.95, and 27.8 → 19.8 minutes.
Some deliverables carry dates the models invented. One brief, the postmortem, scored lower on the newer build (0.95 → 0.85).
MEASUREDDGX Spark · Aug–Sep 2026
Real nights
A month of real jobs read from one DGX Spark’s job log, the source of the hardware card and the prefill and decode figures below.
- Jobs
- 361
- Landed attempts
- 5,361
- App turns served locally
- 98%
Races of 250–300 attempts found 25–37 distinct solution clusters and were still finding new ones in their last 50 attempts. Cloud spend on the published night: $0.00.
Judging is the weak point: 8% of shortlisted candidates were scored, and 92 of 104 outcomes are not yet decided. Interactive latency: median 36.8 s.
