---
license: apache-2.0
base_model: Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2-GGUF
library_name: llama.cpp
pipeline_tag: text-generation
tags:
  - qwen
  - qwen3.5
  - gguf
  - ollama
  - medical
  - russian
  - text-generation
language:
  - ru
  - en
---

# qwen35-9b-medical

`qwen35-9b-medical` is an Ollama/GGUF medical assistant profile based on
`Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2-GGUF`, distributed
locally through the Ollama model
`kwangsuklee/Qwen3.5-9B.Q4_K_M-Claude-4.6-Opus-Reasoning-Distilled-v2`.

This repository contains a quantized GGUF model plus an Ollama `Modelfile` with
a Russian medical-study system prompt and conservative generation parameters.
It is not a formally validated clinical model and is not a substitute for a
licensed clinician.

## Model Details

- Architecture: Qwen3.5
- Parameters: 9B
- Quantization: Q4_K_M
- Runtime target: Ollama / llama.cpp-compatible GGUF runtimes
- Context configured in this profile: 8192 tokens
- Primary language target: Russian medical education
- Secondary language support: English prompts
- Profile type: Ollama instruction/profile customization, not a new full-weight
  supervised fine-tune

## Intended Use

This model profile is intended for medical study and structured clinical
reasoning practice:

- pediatrics learning;
- pathogenesis, symptoms, diagnostics, and treatment principles;
- differential diagnosis practice;
- red flags and safety-oriented next steps;
- explaining medical concepts in clear Russian;
- handling incomplete clinical information more carefully.

## Out-of-Scope Use

Do not use this model as the only basis for:

- diagnosis;
- treatment decisions;
- medication dosing;
- emergency triage;
- replacing a physician or local clinical protocol;
- processing private patient data without appropriate consent and safeguards.

## Safety Behavior

The Ollama profile instructs the model to:

- avoid inventing diagnoses, dosages, studies, percentages, and clinical
  criteria;
- state when there is not enough information;
- separate facts, hypotheses, likely causes, dangerous conditions, and red
  flags;
- consider pediatric factors such as age, body weight, development,
  contraindications, and safety;
- suppress visible reasoning traces such as `<think>` blocks.

## Files

- `qwen35-9b-medical.Q4_K_M.gguf` - quantized GGUF weights.
- `Modelfile` - Ollama profile with template, system prompt, and parameters.
- `USAGE.md` - local usage instructions.
- `MODEL_CARD_RU.md` - Russian model card and safety notes.

## Ollama Usage

Download this repository, then create a local Ollama model from the included
`Modelfile`:

```bash
ollama create qwen35-9b-medical -f Modelfile
ollama run qwen35-9b-medical
```

Example prompt:

```text
Ребенок 5 лет: температура 38.8, кашель, одышка, втяжения межреберий.
Разбери по схеме: вероятно, что опасно исключить, диагностика, красные флаги,
что делать сейчас. Без дозировок.
```

## Generation Parameters

```text
temperature: 0.25
top_p: 0.85
top_k: 40
repeat_penalty: 1.1
presence_penalty: 0.2
num_ctx: 8192
```

## Local Smoke Benchmark

A small local comparison against the base profile was run on four medical
prompts. The benchmark checked visible thinking leakage, uncertainty handling,
red-flag behavior, and answer structure. This is a smoke test only, not a
clinical validation benchmark.

| Metric | Base model | Medical profile |
|---|---:|---:|
| Visible thinking leakage | 2/4 prompts | 0/4 prompts |
| Average response time | 31.0s | 36.1s |
| Medical answer structure | weaker | stronger |
| Missing-data handling | weaker | stronger |
| Red-flag behavior | mixed | better |
| Clinical factual reliability | not formally validated | not formally validated |

## Limitations

- The model can hallucinate.
- It can provide incomplete or outdated medical information.
- It can overgeneralize if the prompt lacks age, weight, symptom duration,
  exam findings, comorbidities, or medication history.
- Current guidelines and medication information must be verified from
  authoritative medical sources.
- The model card does not claim formal clinical validation.

## License

The uploaded artifact is based on the upstream model named above. Reuse must
comply with the Apache-2.0 license and any applicable third-party terms from the
upstream model lineage.