Jeff-Qwen3.5-2B

Use it to make fast, calibrated choices such as ticket triage, moderation, tool selection, or contract-clause classification, with an adapter suited to the task.

Availability

Base and adapter checkpoints available on Hugging Face at revision v1.3; GGUF files available for llama.cpp.

Pricing model

Not established

Deployment

Can be served locally; instructions cover NVIDIA GPUs and Apple silicon.

Privacy

No cloud GPUs used for base-model training

Licence

Licences differ by adapter; sanctions and soc are CC BY-NC 4.0, while aml and trading-desk have data-source terms.

• The page describes v1.2 as a community preview and links to its Hugging Face model repository.
• The model returns a calibrated probability for each option from a single forward pass.
• The Qwen models support choice questions with up to 254 options. ^[1]

  question types: `choice` (pick one of up to 254 options with the Qwen models, 26 with Jeff-Gemma4-E2B), `noul` (yes/no, returned as a probability) and `score` (a

• The v1.2 benchmark-panel score is 78.7% across five benchmarks.
• The model takes about 22 ms per decision on an RTX PRO 6000, according to the page.
• Nine LoRA adapters are available for v1.2 and can be selected by name in a request.
• Jeff returns a calibrated probability for each listed option from a single forward pass. ^[2]

  You describe a situation and list the options in plain words; Jeff returns a calibrated probability for each option

• The v1.3 base and adapters are available as Hugging Face downloads, with GGUF versions also provided. ^[2]

  Weights: `mstrasser/jeff-adapter-<name>` on Hugging Face (revision `v1.3`), with the GGUF versions in

• The quick start supports NVIDIA GPUs and Apple silicon. ^[2]
• The announcement lists 15 adapters for task-specific decisions. ^[2]

  for name in guard tools ground; do # the adapters you want, one folder each

• The source recommends v1.2 for zero-shot use without an adapter. ^[2]

  For zero-shot use without an adapter, use v1.2.