edtorre/ hermes-frenqwen:latest

12 Downloads Updated 3 days ago

The famous qwen3.8:27b optimized for use in agent harnesses like Hermes Agent. Recommended for 24gb VRAM + systems. Its about 21gb in size.

vision tools thinking

ollama run edtorre/hermes-frenqwen

curl http://localhost:11434/api/chat \
  -d '{
    "model": "edtorre/hermes-frenqwen",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='edtorre/hermes-frenqwen',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'edtorre/hermes-frenqwen',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 3 days ago

3 days ago

8724e82c4d3d · 18GB

model

archqwen35

·

parameters27.3B

·

quantizationQ4_K_M

17GB

projector

archclip

·

parameters461M

·

quantizationBF16

931MB

license

Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US

11kB

params

{ "min_p": 0, "num_ctx": 65536, "presence_penalty": 0, "repeat_penalty": 1, "tem

108B

Readme

Hermes FrenQwen 🐑⚡

Qwen3.8-27B tuned for agent work in the Hermes Agent harness — 64k+ context baked in, vendor-spec sampling pinned, zero system-prompt interference.

ollama run edtorre/hermes-frenqwen

What’s different from the base model?

Setting	Value	Why
Context window	65,536	Hermes requires ≥64k for multi-step tool use. Native 262k, so no YaRN needed
Sampling	`temp 1.0 · top_p 0.95 · top_k 20 · min_p 0`	Qwen team’s published thinking-mode spec — avoids repetition collapse in reasoning models
Output limit	unlimited	Thinking traces are never truncated before a tool call completes
System prompt	none	Hermes injects its own (memory, skills, tool schemas)

Capabilities inherited from qwen3.8:27b-q4_K_M: vision · tool calling · thinking (levels false / low / medium / xhigh, default medium).

Hardware note

~21 GB loaded at full context (weights + KV). Runs 100% on 3× RTX 3060 (12 GB each); should also fit a single RTX 4090 or a Mac with ≥32 GB unified memory.

Sample chat

curl http://localhost:11434/api/chat -d '{
  "model": "edtorre/hermes-frenqwen",
  "messages": [{"role": "user", "content": "Hello!"}]
}'

License: Apache 2.0 (weights © Qwen team)

# Hermes FrenQwen 🐑⚡

Qwen3.8-27B tuned for agent work in the **Hermes Agent harness** — 64k+ context baked in, vendor-spec sampling pinned, zero system-prompt interference.

```
ollama run edtorre/hermes-frenqwen
```

## What's different from the base model?

| Setting | Value | Why |
|---|---|---|
| Context window | **65,536** | Hermes requires ≥64k for multi-step tool use. Native 262k, so no YaRN needed |
| Sampling | `temp 1.0 · top_p 0.95 · top_k 20 · min_p 0` | Qwen team's published thinking-mode spec — avoids repetition collapse in reasoning models |
| Output limit | unlimited | Thinking traces are never truncated before a tool call completes |
| System prompt | none | Hermes injects its own (memory, skills, tool schemas) |

Capabilities inherited from `qwen3.8:27b-q4_K_M`: **vision · tool calling · thinking** (levels `false / low / medium / xhigh`, default `medium`).

## Hardware note

~21 GB loaded at full context (weights + KV). Runs 100% on 3× RTX 3060 (12 GB each); should also fit a single RTX 4090 or a Mac with ≥32 GB unified memory.

## Sample chat

```bash
curl http://localhost:11434/api/chat -d '{
  "model": "edtorre/hermes-frenqwen",
  "messages": [{"role": "user", "content": "Hello!"}]
}'
```

License: Apache 2.0 (weights © Qwen team)

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)