Gemma-3-27B
GemmaDense transformer with sliding-window attention
layers 62
d_model 4608
params 27B
released 2025-Q4
Probes for this model
0No probes registered yet — open territory.
3 models registered. Probes are model-specific by design — a Qwen3.6 probe will not transfer to Llama-3 without re-training. Cross-model transfer numbers are reported via Pearson_CE on each probe DNA page.
Cell value = number of registered probes for that model × category combination. Empty dashed cells indicate categories where no probe has been registered yet for that model — these are the highest-leverage targets for new submissions.
Dense transformer with sliding-window attention
No probes registered yet — open territory.
Dense transformer (instruct-tuned)
Hybrid GDN + Gated-Attn (dense, reasoning)
These are the models the open-weights community can fork, fine-tune, redistribute. ProbeBench prioritizes coverage here.
Llama, Gemma, others. Subject to original license — research use generally OK; commercial use varies.
We accept closed-weight probes (e.g. GPT-4) but cap their license score at 0.5 × 0.05 = 0.025.
Hybrid architectures (Qwen3.6 GDN, Mamba SSMs, MoE) require model-specific probe-extraction code. The openinterp SDK auto-detects layer paths via model.language_model.layers[N] for HF transformers and model.layers[N] for dense paths. Probes for hybrid models declare the position field carefully — token_avg vs end_question vs mid_think have very different semantics on reasoning models.
Have a model that should be on here? PR a registry entry against lib/probebench-data.ts. Required fields follow the ModelEntry schema in lib/probebench-types.ts.
id: "Qwen/Qwen3.6-27B" short_name: "Qwen3.6-27B" family: "Qwen" param_count: "27B" architecture: "Hybrid GDN + Gated-Attn" layers: 64 d_model: 5120 release: "2026-04" weights_license: "Apache-2.0" hf_url: "https://huggingface.co/Qwen/Qwen3.6-27B" thinking_mode: true