Model registry
Sixteen ranked arms, one adjudicator. Every row carries its lane, its licence, its commercial-use status and what will actually run when you press go.
Live
0
provider key present and a client wired in this codebase
Simulated profile
15
runs a published-benchmark prior, labelled everywhere it appears
Needs key
5
client wired, awaiting OPENROUTER_API_KEY, REPLICATE_API_TOKEN, ROBOFLOW_API_KEY
Not pursued
1
a recorded decision, kept visible on purpose
The bench
Rank order is the research ranking, not a result. Results live on the evidence page.
| # | Model | Provider | Licence | Status | Why it is on the list |
|---|---|---|---|---|---|
| 1 | RF-DETR Small/Medium (+SAHI) Supervisedrf-detr | Roboflow | Apache-2.0 | SIMULATED PROFILE Simulated — ROBOFLOW_API_KEY is not set. | Best domain transfer per label of any 2026 detector (DINOv2 backbone), Apache-clean, and the safest bet to win the table on 5–10 m imagery. Roboflow serverless: 15k free inferences, then credits; US$99/mo Core needed to keep Keith's imagery private. |
| 2 | SAM 3.1 (exemplar + concept) Exemplarsam3 | Roboflow | SAM License (commercial use permitted, carve-outs + patent clause) COMMERCIAL: CONDITIONAL | SIMULATED PROFILE Simulated — ROBOFLOW_API_KEY is not set. | Consumes Keith's click-box directly and returns EVERY matching instance with a mask and an ID in one pass — zero training, and fine-tuning code ships. Zero-shot/exemplar SAM 3 runs on Roboflow serverless; fine-tuned SAM 3 needs a dedicated deployment. |
| 3 | YOLO26-s/m + P2 head (1280–1536, SAHI) Supervisedyolo26-p2 | Replicate | AGPL-3.0 (or paid Enterprise) COMMERCIAL: CONDITIONAL | SIMULATED PROFILE Simulated — hosted on Replicate but not wired here; AGPL must be resolved before delivery. | The reference workhorse: STAL small-target label assignment plus a stride-4 P2 head is the best-understood recipe for 5–15 px objects. ⚠ AGPL: hosted delivery triggers source disclosure. Resolve the licence before this arm ships to a client. |
| 4 | GeCo2 Exemplargeco2 | Local weights | Unverified — check repo LICENSE before shipping COMMERCIAL: CONDITIONAL | SIMULATED PROFILE Simulated — local weights only (GitHub jerpelhan/GECO2) — no hosted API in the four-key plan. | AAAI 2026 SOTA few-shot counting-by-detection, engineered for scale generalisation — the exact failure mode at 5–10 px (22–27% RMSE below GeCo). Local weights + Detectron2; no hosted API. Free to run, costs a GPU and an afternoon. |
| 5 | CountGD Exemplarcountgd | Hugging Face | MIT | SIMULATED PROFILE Simulated — Hugging Face Space / local build — no hosted API in the four-key plan. | The only permissively-licensed exemplar counter with a live public demo — a real count on the first September images before any training exists. HF Space (ZeroGPU) free; PRO US$9/mo for 8× quota. Local build is Python 3.9 / GCC 11, budget half a day. |
| 6 | T-Rex2 (interactive visual prompt) Exemplart-rex2 | DeepDataSpace | IDEA License 1.0 (API-gated) COMMERCIAL: CONDITIONAL | SIMULATED PROFILE Simulated — DeepDataSpace account not opened (same vendor as DINO-X, decision G3). | The model that invented the draw-one-box-find-them-all workflow Keith described; embedding reuse across a whole flight is the UX argument in one call. ~US$0.014/call on DeepDataSpace. Same account as DINO-X, which G3 did not open. |
| 7 | DINO-X Pro Zero-shotdino-x | DeepDataSpace | Closed commercial API COMMERCIAL: CONDITIONAL | NOT PURSUED Not pursued (G3): DeepDataSpace account deliberately not opened — payment friction plus a data-residency question over consent imagery. | Best long-tail open-world model of 2026 (+5.8 AP on LVIS-rare) with native visual-prompt counting — the zero-training upper bound we chose not to buy. NOT PURSUED (G3). ~0.1–0.15 CNY/call, WeChat-Pay friction, China data residency for consent imagery. |
| 8 | Gemini Flash (tiled) VLMgemini-flash | OpenRouter | Closed commercial API | SIMULATED PROFILE Simulated — OPENROUTER_API_KEY is not set. | Joint-top of Roboflow's counting leaderboard (82.4%) and the cheapest credible frontier arm — but ViT patches are 14–16 px, so watch it die below 8 px. ~US$0.001 per 1024 px tile via OpenRouter (+5.5% credit fee). 500 images ≈ US$3. |
| 9 | DEIM-D-FINE-L Superviseddeim-d-fine | Local weights | Apache-2.0 | SIMULATED PROFILE Simulated — self-hosted Apache weights — no hosted endpoint. | Best Apache-licensed accuracy/speed point in 2026 (54.7 AP @ 124 FPS T4) — the drop-in if the AGPL question kills YOLO26. Self-hosted; ~50% less training time than the comparable YOLO recipe. |
| 10 | YOLOE-26 (visual + prompt-free) Zero-shotyoloe-26 | Replicate | AGPL-3.0 (or Enterprise) COMMERCIAL: CONDITIONAL | SIMULATED PROFILE Simulated — Replicate hosts a YOLOE variant but it is not wired here. | Text, visual AND prompt-free inference — the only arm that can find things nobody thought to ask for. Licence is the catch. Replicate hosts a YOLOE variant at near-zero cost per run; AGPL applies to delivery. |
| 11 | Qwen3-VL-32B-Instruct (self-hosted) VLMqwen3-vl | Local weights | Apache-2.0 | SIMULATED PROFILE Simulated — self-hosted open weights — that IS the point of this arm. | The only frontier-class counter that runs entirely on Keith's own hardware — the direct answer to 'cloud is painful for our datasets'. Open weights, no API bill; needs a serious GPU. Local-weights arm by design. |
| 12 | OWLv2 (image-conditioned one-shot) Zero-shotowlv2 | Replicate | Apache-2.0 | SIMULATED PROFILE Simulated — not on Roboflow serverless; Replicate route not wired here. | Permissive one-shot detection from a single crop — the licence-safe fallback for the whole exemplar UX if SAM 3 and T-Rex2 both hit legal friction. Not on Roboflow serverless; reachable via Replicate (owl-sam / owlvit) at ~US$0.001/run. |
| 13 | LOCA + DAVE (density counting) Exemplarloca-dave | Local weights | Unverified — check repo LICENSE before shipping COMMERCIAL: CONDITIONAL | SIMULATED PROFILE Simulated — local weights only (ViCoS Lab) — no hosted API. | The dense-mat insurance policy: when a good year produces mats and every box detector saturates, density regression is the only defensible number. Local weights (ViCoS Lab). Density lane — exposed under the exemplar lane because the contract has no density lane. |
| 14 | Grounding DINO (open Swin-B) Zero-shotgrounding-dino | Replicate | Apache-2.0 | SIMULATED PROFILE Simulated — REPLICATE_API_TOKEN is not set. | Free, self-hostable, fine-tunable, and the backbone CountGD sits on — its job here is to document the honest zero-shot text-prompt floor. ~US$0.0014/run on Replicate (adirik/grounding-dino). Prepaid credit is a hard spend cap. |
| 15 | RTMDet-s / RTMDet-tiny Supervisedrtmdet | Local weights | MIT | SIMULATED PROFILE Simulated — self-hosted MMDetection weights — the desktop-throughput arm. | 300–1000+ FPS on one card: the arm that makes 3,000 images per hectare tractable on Keith's desktop instead of a cloud bill. Self-hosted MMDetection. Throughput specialist — accuracy is not the reason it is here. |
| 16 | GPT-5.6 Sol (tiled) VLMgpt-5-6 | OpenRouter | Closed commercial API | SIMULATED PROFILE Simulated — OPENROUTER_API_KEY is not set. | One more frontier arm for completeness — behind Gemini and Qwen on both detection and counting, at roughly 3× the price. Via OpenRouter; ~3× Gemini Flash per tile. Included for coverage, not expectation. |
The adjudicator lane
Deliberately excluded from the table above — filtered by lane, not by id.
Claude (crop adjudicator)AdjudicatorLIVEclaude-adjudicator
Not a counter: a second opinion on 512 px crops at 1:1 pixel mapping, and a tile-triage pass. Hidden from the leaderboard by lane.
It is a referee, not a counter: it sees 512 px crops at 1:1 pixel mapping and answers is this a plant or a stone, and it does a cheap tile-triage pass so the expensive arms only look at tiles worth looking at. It never produces a count for the leaderboard, because a model that adjudicates the evidence cannot also be scored on it.
Licence: Closed commercial API. Haiku-class per crop; Batch API halves it. Never used for whole-image counting.
Why some rows say SIMULATED PROFILE
Access is computed from the environment at request time, never asserted in a config file. An arm is LIVE only when its provider key is present and this codebase has a client wired to it. Everything else runs a deterministic error profile derived from that model’s published benchmarks — which is a defensible way to demonstrate the harness before September, and is not a measurement. Simulated results are labelled on the leaderboard, in the run manifest and in every export, and each one is replaced by a real number the moment that arm runs live. No simulated figure is ever presented as a live result.Decision G3 — DINO-X not pursued
DINO-X is the strongest zero-shot model on this list and we are consciously not opening the account: China-hosted API plus WeChat-Pay friction, which is a data-residency question we are not willing to answer on someone else’s consent imagery. SAM 3 and CountGD carry the exemplar lane without it, and T-Rex2 sits behind the same vendor account for the same reason. The row stays visible with its reason attached — a missing row looks like an oversight, a “not pursued” row looks like a decision.Licences travel with the model
AGPL-3.0 on the YOLO arms is a delivery problem, not a footnote: hosted delivery triggers source disclosure, so that licence must be resolved (or the arm swapped for the Apache-licensed DEIM-D-FINE) before anything ships to a client. Rows marked COMMERCIAL: CONDITIONAL have a real condition attached — read the licence column before quoting a result from one.