Field notes from local models actually in use. What the measurements say, what the instrument got wrong, and which layer was carrying the result.
Nothing here ranks a model until the instrument that measured it has been checked.
Personal benchmarking on hardware I actually use, published with the configuration attached. Each page carries its own date and moves when the numbers move.
RTX 3090 against a Ryzen AI MAX 390 on the same blobs. Which model holds the seat on each chip, and why decode rate does not decide it.
Read the numbers →Interactive. Drag the context and watch layers spill to the CPU. Three of sixty-six costs 27.6%. Flip decode to wall clock and the ranking inverts.
Open the model →Same architecture, new weights. Speed, multi-token prediction on the 3090, and correctness across the same three fixtures.
Read the A/B →Five-sweep intervals for synthetic prefill across three configurations. Repeatability evidence, not a quality ranking.
Read the study →Five matched prompts per service across TestBench, Z13, and hosted Luna. Client-observed latency, not a hardware ranking.
Read the study →A single staged GPT-6 Luna observation over local Watch/Zepp tools. Strict and semantic scores are both shown; not a model ranking.
Read the case study →Measured 28-29 Sep 2026 on two RTX 3090s in a headless test bench, one card against both. Most of the roster fits one 24 GB card and gains nothing from a second. Two models do not fit, and they behave differently.
| Model | Size | One card | Two cards |
|---|---|---|---|
| qwen3-next (80B MoE) | 50 GB | 46% in VRAM, 25.6 tok/s | 92-96% in VRAM, 72.6 tok/s |
| nemotron-3.5-lightning (30B MoE) | 25.4 GB | 87.5% in VRAM, 49-58 tok/s | 100% in VRAM, 75-88 tok/s |
qwen3-next needs both cards. On one it runs at about a third of the speed with more than half the model on the CPU. nemotron-3.5-lightning is borderline: its weights nearly fit one card, but with its context and working memory about an eighth spills to the CPU. It still runs there at roughly two-thirds of its two-card speed, and holds that speed as context fills (75 tok/s at 115k tokens on two cards).
On the same five repair tasks, with every sampling parameter pinned identical, nemotron-3.5-lightning, qwen3-next Instruct and qwen3.6:35b scored 30, 27 and 29 of 30. At six trials per task that is a tie, not a ranking. Where they differ is wall clock: 8.6, 25.0 and 5.7 minutes for the thirty cells.
qwen3-next Thinking, run as designed with reasoning on (32k context, temperature 0.6), scored 29 of 30 with no timeouts, in 69.6 minutes. It is reported on its own because its settings differ from the three above.
Earlier qwen3-next results on this site used the Thinking variant run in non-thinking mode. They describe that configuration, not the model.
Two pages about measurement rather than models. Both exist because a number looked like a finding and was not.
Eligibility, roster history, and prompt-level screening evidence. Historical runs are distinct from the current specification; the page keeps their scope and limits separate.
See the screening →How E9, Fusion, hardware performance, and interactive latency relate without becoming one ranking; protocols, screening status, and what may never be pooled.
Read the protocol →Which layer was actually carrying the result, and what changes when you swap the harness instead of the model.
Read the layer →Thirty-seven measurement faults, and the finding each one would have been published as. Seventeen in the first week, twenty since. None of them were the model.
Read the count →Papers 1.16 and 1.19. Subprocess capture embedded terminal sequences that corrupted multi-line JSON, and a model took the blame for it. Under clean REST capture it passes all six probes.
The terminal failed, not the model →Do not ask only which model is strongest. Ask which layer is carrying the result: grounding, routing, protocol discipline, repair policy, role fit, or raw model power.
The framework material that used to sit on this page has its own home now, because it is not about local models.