Skip to content

GLiNER2.5 multi (v1 + classifier)

  • Status: m0 (verified end-to-end)
  • Adapter: gliner2 — split encoder + classifier, schema with [P] / [C] / [DESCRIPTION] / [SEP_TEXT]
  • License: Apache-2.0
  • Parameters: 287M · Architecture: GLiNER2.5 multi-v1 · Languages: multilingual

Default variant
Repo jugaadsrl/gliner2.5-multi-v1-onnx
Files fp16_25/encoder_fp16.onnx, fp32_25/classifier_fp32.onnx, fp16_25/tokenizer.json
Quantization fp16 encoder + fp32 classifier
Download 577 MB
Role default; fp32 is a quality variant

Variants

Variant Quantization Size Role Install
fp16 * fp16 encoder + fp32 classifier 577.1 MB default dtk models download gliner2.5-multi --variant fp16
fp32 fp32 encoder + fp32 classifier 1131.9 MB quality dtk models download gliner2.5-multi --variant fp32

dtk models download gliner2.5-multi gets fp16; --variant fp32 downloads the quality build alongside it. dtk models use gliner2.5-multi --variant fp32 makes it persistent. See Variants and quantization.

Reproducibility

  • Revision: 3de2c11f55853b7919f985a59bfc8df27b5ec991
  • Default file hashes (SHA256):
  • fp16_25/encoder_fp16.onnx — b389910b27261d4999bd41c164bfa0334f76b3bc9fd3d890527b6f07fa1bfb94
  • fp32_25/classifier_fp32.onnx — 3c06934244a4763a7411d9dceb002ba2d5c84d1bb7f7d66aa64acd5653314454
  • fp16_25/tokenizer.json — c62446df87ae18ec98b133f8f84fc449a07cc89bbf8ef192a4cb5f9c53777a7a
  • Verify: dtk models verify gliner2.5-multi --variant fp16
  • Bench: uv run python bench/run.py --suite bench/suites/smoke.jsonl --models gliner2.5-multi
  • The fp32 variant is pinned to the same revision; its hashes live in catalog.py.

Evidence (2026-10-19)

  • 9/15 on the 15-case Spanish routing suite (meeting / tasks / contact / idea / other categories).
  • 353 ms/decision on CPU.
  • Same accuracy as the local int8 build used during development; per-case agreement with that build was not recorded.

Contract notes

  • The classifier input name is read from the graph (choice_states in this export, hidden_state in others), so the adapter does not hardcode it.
  • Labels and descriptions travel in the request; no prompt template is generated.
  • task defaults to classify; override per question with a raw task field.

Known limitations

  • This is the base multi-v1 export, not the multi-Decide fine-tune (no compatible export published yet).
  • English/multilingual tokenizer fallback: the run above used the raw tokenizers fallback; a transformers tokenizer may change scores slightly.
  • Benchmarked on a synthetic Spanish routing set; validate on your own data.

Usage

dtk models download gliner2.5-multi                       # fp16 default
dtk models download gliner2.5-multi --variant fp32        # quality variant
dtk decide --model gliner2.5-multi --state "Meeting with Acme: 10% discount…" \
    --q cat='choice:meeting|tasks|contact|idea|other'

# use the quality variant for one request
dtk decide --model gliner2.5-multi --variant fp32 \
    --state "Meeting with Acme: 10% discount…" \
    --q cat='choice:meeting|tasks|contact|idea|other'

Provenance

Conversion published by jugaadsrl; base model fastino/gliner2.5-multi-v1.