RefinedNeuro/ vibethinker-3b-hermes:Q4_K_M

584 Downloads Updated 1 week ago

Fine Tuned For Hermes-style function calling with VibeThinker 3b

tools

ollama run RefinedNeuro/vibethinker-3b-hermes:Q4_K_M

curl http://localhost:11434/api/chat \
  -d '{
    "model": "RefinedNeuro/vibethinker-3b-hermes:Q4_K_M",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='RefinedNeuro/vibethinker-3b-hermes:Q4_K_M',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'RefinedNeuro/vibethinker-3b-hermes:Q4_K_M',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 1 week ago

1 week ago

6fc56827f0c5 · 2.5GB ·

model

archqwen2

·

parameters3.09B

·

quantizationQ6_K

2.5GB

template

{{- if or .System .Tools }}<|im_start|>system {{- if .System }} {{ .System }} {{- end }} {{- if .Too

1.2kB

params

{ "repeat_penalty": 1.1, "stop": [ "<|im_end|>", "<|endoftext|>" ],

112B

Readme

refinedneuro/vibethinker-3b-hermes

A 3B reasoning model that calls tools. A LoRA fine-tune of WeiboAI/VibeThinker-3B (Qwen2-based) that adds Hermes-style function calling (<think>…</think> + <tool_call>…</tool_call>) while preserving its math/STEM reasoning.

Run

ollama run refinedneuro/vibethinker-3b-hermes

Tags

latest (=Q6_K), Q6_K, Q8_0, f16

Quant note. This is a tool-calling model, and quantization below Q6_K degrades tool-call fidelity (Q3/Q4/Q5 emit malformed/incomplete calls and can loop). For that reason the Q3_K_M / Q4_K_M / Q5_K_M tags now point to the Q6_K build — pulling any of them gives you the smallest good-quality model. Use Q6_K or higher; default latest = Q6_K.

Recommended settings (baked in)

temperature 0.6 · top_p 0.95 · repeat_penalty 1.1 · stops on <|im_end|>. For reliable single tool calls, also stop on </tool_call>.

Benchmarks (summary)

Reasoning (AIME 2024): base avg@4 0.842 → this model 0.783; pass@4 0.867 unchanged.
Tool-calling (BFCL single-turn, with stop fix): simple ~60%, multiple ~31%, relevance 87.5%.

Known limitations

Over-eager to call tools; can fall into repetition loops on out-of-distribution multi-turn input. Not production-ready. See the full model card: https://huggingface.co/RefinedNeuro/VibeThinker-3B-Hermes

License

Apache-2.0 — built on VibeThinker-3B and lambda/hermes-agent-reasoning-traces (both Apache-2.0).

# refinedneuro/vibethinker-3b-hermes

A 3B **reasoning model that calls tools**. A LoRA fine-tune of
[WeiboAI/VibeThinker-3B](https://huggingface.co/WeiboAI/VibeThinker-3B) (Qwen2-based) that
adds **Hermes-style function calling** (`<think>…</think>` + `<tool_call>…</tool_call>`)
while preserving its math/STEM reasoning.

## Run
```bash
ollama run refinedneuro/vibethinker-3b-hermes
```

## Tags
`latest` (=Q6_K), `Q6_K`, `Q8_0`, `f16`

> **Quant note.** This is a tool-calling model, and quantization below **Q6_K** degrades
> tool-call fidelity (Q3/Q4/Q5 emit malformed/incomplete calls and can loop). For that
> reason the **`Q3_K_M` / `Q4_K_M` / `Q5_K_M` tags now point to the Q6_K build** — pulling any
> of them gives you the smallest *good-quality* model. Use **`Q6_K` or higher**; default
> `latest` = Q6_K.

## Recommended settings (baked in)
temperature 0.6 · top_p 0.95 · repeat_penalty 1.1 · stops on `<|im_end|>`.
For reliable single tool calls, also **stop on `</tool_call>`**.

## Benchmarks (summary)
- Reasoning (AIME 2024): base avg@4 0.842 → this model 0.783; pass@4 0.867 unchanged.
- Tool-calling (BFCL single-turn, with stop fix): simple ~60%, multiple ~31%, relevance 87.5%.

## Known limitations
Over-eager to call tools; can fall into repetition loops on out-of-distribution multi-turn
input. **Not production-ready.** See the full model card:
https://huggingface.co/RefinedNeuro/VibeThinker-3B-Hermes

## License
Apache-2.0 — built on VibeThinker-3B and lambda/hermes-agent-reasoning-traces (both Apache-2.0).

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