Open-weight medical language model

Pestle-27B Ternary

8.2× smaller text weights with 96.48% median capability retention. Built for private local and on-premises systems.

ReleasedApache 2.0
Pestle27B · ternary medical model
6.75 GBDeployed text weights
2.52Effective BPW
89.79MedQA
89.02HumanEval pass@1
89.79MedQA · full test split
86.89MMLU medical subjects
75.28BioASQ · token F1
90.15PharmaRAG · MRR
01 · The model

Specialist capability without a permanent cloud dependency.

Pestle-27B-Ternary is a medical-focused language model derived from the Qwen3.6-27B architecture and compressed for local serving. It is a public proof point for a broader Doses AI capability: making specialist models smaller, private, and practical to operate inside the institution that owns them.

Runnable release

8.2× smaller deployed text weights

Pestle carries 6.75 GB of deployed text weights, compared with 55.56 GB for the source Qwen3.6-27B FP16 model. The complete shipping package is one 8.48 GB runnable GGUF.

55.56 GB
6.75 GB
Release posture

Weights, runtime, protocol, evidence.

The release pairs the runnable model with the Mortar inference runtime, deterministic generation settings, full-suite benchmark summaries, and retained evaluation evidence. Developers can inspect the operating point rather than relying on a single leaderboard number.

Hugging Face ↗ Mortar runtime ↗
02 · Evidence

Measured across clinical knowledge, literature and retrieval.

Pestle is evaluated across multiple medical task families, from USMLE-style reasoning to open-ended biomedical evidence extraction and pharmaceutical retrieval.

Selected full-suite results
89.79
86.89
90.15
75.28
76.70
68.85
BenchmarkMetricScore
MedQAAccuracy89.79
MMLU medicalAggregate86.89
BioASQExact match55.94
PharmaRAGnDCG@1084.62
PubMedQAMacro F162.78
ChemBenchScore61.72
MedXpertQAAccuracy32.49
HumanEvalpass@189.02

Scores and protocols follow the retained release record. See the model card for task definitions, row counts, generation settings, comparison provenance, and full evidence paths.

03 · In use

One local model. Several narrow, reviewable workflows.

Pestle is designed as a building block for assistive systems whose sources, prompts and outputs remain under the operator's control.

+

Clinical reasoning support

Draft explanations over structured medical questions and institution-approved knowledge, with professional review.

Biomedical evidence

Extract grounded spans and concise answers from supplied abstracts, policies and literature.

Private retrieval

Generate answers beside local formularies, guidelines or pharmaceutical evidence stores.

{ }

Health-tech engineering

Support coding and structured-output workflows without routing proprietary context to a public model API.

04 · Deployment

Keep the model next to the data.

Local deployment removes external inference as a structural dependency, letting organisations design access, retrieval, logging and review inside their own governance boundary.

Private sourcesGuidelines, formularies, literature, policies and approved records.
Pestle + Mortar6.75 GB of text weights in an 8.48 GB local GGUF, with an OpenAI-compatible serving path.
Reviewed outputStaff-facing drafts and evidence with human oversight and local audit.
Hospitals & health systems

Compress models already adapted to the institution.

Doses AI can work with specialist checkpoints and local acceptance suites to produce smaller deployment artifacts without making a third-party inference endpoint mandatory.

Pharma & life sciences

Keep therapeutic-area models and evidence private.

Compress medical-information, literature, safety, regulatory or retrieval models for operation near proprietary corpora and governed internal workflows.

Research model, not a medical device. Pestle must not directly determine diagnosis, treatment, prescribing, patient management or emergency triage. Outputs require qualified professional review. Benchmark performance does not establish clinical safety or efficacy.

Build private medical AI on hardware you control.

Start a conversation