Local Hotel Source Selection: 28-Model Benchmark

.ob-table{border-collapse:collapse;width:100%;font-size:14px}.ob-table th,.ob-table td{border:1px solid #d8dee4;padding:7px;vertical-align:top}.ob-table th{background:#f6f8fa;text-align:left}.ok{background:#e9f8ee}.warn{background:#fff8dc}.bad{background:#fdecec}.mono{font-family:ui-monospace,SFMono-Regular,Menlo,monospace}.note{border-left:4px solid #4b74c6;padding:10px 14px;background:#f5f8ff}.metric{display:inline-block;padding:7px 10px;margin:3px;background:#f2f4f7;border-radius:5px}pre{white-space:pre-wrap;overflow-wrap:anywhere;font-size:12px}details{margin:10px 0}.small{font-size:13px;color:#4b5563}

Scope: This is a source-selection benchmark, not an end-to-end browsing benchmark. Every model received the same frozen 16-result candidate set produced by deterministic search and had to choose the property-specific official and guest-review corpus. That isolates the judgment task a reliable acquisition controller should delegate to a model; it does not claim these models can operate a browser reliably.

We tested all 28 installed general-purpose text-model tags available on the Mac mini and the remote worker: 19 local and 9 remote. Coding, vision, and embedding models were excluded. Each physical host ran one model at a time; local and remote queues ran concurrently.

28 runs18 schema-valid13 exact8 wrong contract2 request failures

Result grid

Model Machine Status Score / 100 Wall seconds What it found
gemma4:e4b remote-10.7.6.39 accepted 100.00 14.10 Exact expected corpus
qwen3:8b local-mac-mini accepted 100.00 22.60 Exact expected corpus
muse-glimmer:30b remote-10.7.6.39 accepted 100.00 23.45 Exact expected corpus
qwen3.8:latest remote-10.7.6.39 accepted 100.00 25.69 Exact expected corpus
edtorre/muse-glimmer-agent:latest remote-10.7.6.39 accepted 100.00 26.21 Exact expected corpus
qwen3.6:latest remote-10.7.6.39 accepted 100.00 27.54 Exact expected corpus
qwen3.5:35b remote-10.7.6.39 accepted 100.00 28.31 Exact expected corpus
qwen3:14b local-mac-mini accepted 100.00 47.28 Exact expected corpus
gemma3:12b local-mac-mini accepted 100.00 47.58 Exact expected corpus
gemma4:31b remote-10.7.6.39 accepted 100.00 59.84 Exact expected corpus
mistral-small:24b local-mac-mini accepted 100.00 72.18 Exact expected corpus
qwen3.6:latest local-mac-mini accepted 100.00 91.95 Exact expected corpus
qwen3.6:27b local-mac-mini accepted 100.00 105.83 Exact expected corpus
qwen3.5:9b remote-10.7.6.39 accepted 95.00 11.81 missed C04
qwen3.5:9b-64k local-mac-mini accepted 95.00 31.63 missed C04
qwen3.5:9b local-mac-mini accepted 95.00 33.54 missed C04
mo-shakib/clearwriter:latest local-mac-mini accepted 94.86 16.38 missed C06
nemotron-3.5-lightning:30b remote-10.7.6.39 accepted 94.00 36.50 missed C06, C14
nemotron-3.5-lightning:30b-mlx local-mac-mini request_error 0.00 0.04 Request failed
muse-glimmer:30b-mlx local-mac-mini request_error 0.00 0.06 Request failed
gemma4:e2b-mlx local-mac-mini invalid_output 0.00 5.48 Wrong output contract
gemma4:e4b-mlx local-mac-mini invalid_output 0.00 10.23 Wrong output contract
gemma4:26b-16k local-mac-mini invalid_output 0.00 19.58 Wrong output contract
gemma4:26b-mlx local-mac-mini invalid_output 0.00 19.92 Wrong output contract
gpt-oss:20b local-mac-mini invalid_output 0.00 20.47 Wrong output contract
gemma4:12b-mlx local-mac-mini invalid_output 0.00 41.60 Wrong output contract
gemma4:12b-mlx-chat local-mac-mini invalid_output 0.00 44.01 Wrong output contract
qwen3.8:27b-mlx local-mac-mini invalid_output 0.00 62.28 Wrong output contract

Open the full report with the task, candidate catalog, scoring, exact output from every model, and telemetry →

Headline findings

  • 13 models produced the exact expected corpus.
  • The fastest exact result overall was gemma4:e4b on remote-10.7.6.39 at 14.10 seconds.
  • The fastest exact local result was qwen3:8b at 22.60 seconds.
  • The current qwen3.5:9b family correctly chose the canonical page and all five guest sources, but consistently omitted official location page C04, scoring 95.
  • Several Gemma4 MLX tags returned sensible-looking selections under their own invented schemas. They are scored zero because downstream automation needs the agreed contract, not merely readable prose.
  • The remote Nemotron result was schema-valid but selected one non-preferred source; the two local 30B MLX requests failed at the Ollama server before producing output.

Same-model local versus remote speed

We verified identical Ollama digests for qwen3.5:9b and qwen3.6:latest, then ran three paired, simultaneous, cold trials per model with keep_alive=0. Each host ran only one model at a time.

Exact model Local median Remote median Remote speedup Local samples Remote samples
qwen3.5:9b 31.57 10.53 3.00× 33.66, 31.57, 31.55 12.62, 10.51, 10.53
qwen3.6:latest 119.71 29.17 4.10× 94.36, 122.21, 119.71 31.94, 29.17, 27.67

These are observed end-to-end wall times for this prompt, including cold loading. They are not a general hardware benchmark.

What this changes

A larger model is not required for the bounded selection decision. The 8B local Qwen produced the exact corpus in 22.60 seconds, while the existing 9B Qwen missed one useful official page. The correct process remains deterministic search and rendering plus a narrow model selection call. The benchmark does not rehabilitate the original broad 15-turn web agent, which mixed searching, URL guessing, rendering, budgeting, and artifact writing.

Read every model’s exact result →

Similar Posts