beeble/ ministral-3-3b-instruct-2512-fb16:latest

15 Downloads Updated 2 weeks ago

The Ministral 3 family is designed for edge deployment, capable of running on a wide range of hardware. This version is the converted model from Huggingface with full FB16 precision.

ollama run beeble/ministral-3-3b-instruct-2512-fb16

curl http://localhost:11434/api/chat \
  -d '{
    "model": "beeble/ministral-3-3b-instruct-2512-fb16",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='beeble/ministral-3-3b-instruct-2512-fb16',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'beeble/ministral-3-3b-instruct-2512-fb16',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 2 weeks ago

2 weeks ago

d6408f4a6ba6 · 6.9GB

model

archmistral3

·

parameters3.43B

·

quantizationBF16

6.9GB

params

{ "num_ctx": 16384, "temperature": 0.15 }

37B

Readme

Ministral 3 3B Instruct 2512 — FB16 Build

Full-precision FB16 (16-bit float) build of Mistral AI’s Ministral 3 3B Instruct 2512 — no quantization, weights stored exactly as released. Output is (nearly) identical in quality to model version hosted by Mistral itself.

FB16 vs. Standard Q4_0

The default Ollama tag is usually a 4-bit build (q4_0). Key differences:

	FB16 (this build)	Q4_0 (default)
Quality	Bit-exact, zero quantization error	~4.25 bits/weight; on a small 3B model the quality loss is proportionally larger (weaker instruction adherence, more reasoning slips, worse calibration)
Size	~6–7 GB	~2–2.5 GB
Speed	Slower (weights are 4× larger to stream)	Typically 2–3× more tokens/sec

When to use which: Q4_0 for edge hardware, limited RAM/VRAM, and fast casual chat. FB16 when fidelity matters — benchmarking, structured output (JSON, code, tool calls), agentic pipelines where small errors cascade, or generating reference data.

Disclaimer

This build was independently created. The model provider is not acting on behalf of Mistral AI and has no affiliation with Mistral AI whatsoever.
Model weights are subject to Mistral AI’s license — review it before commercial use.

License

Model weights: Mistral AI. This build and README are provided as-is, without warranty.

# Ministral 3 3B Instruct 2512 — FB16 Build

Full-precision **FB16 (16-bit float) build** of Mistral AI's Ministral 3 3B Instruct 2512 — no quantization, weights stored exactly as released. Output is (nearly) identical in quality to model version hosted by Mistral itself.

## FB16 vs. Standard Q4\_0

The default Ollama tag is usually a 4-bit build (`q4_0`). Key differences:

|             | **FB16 (this build)**                    | **Q4\_0 (default)**                                                                                                                                       |
| ----------- | ---------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Quality** | Bit-exact, zero quantization error       | \~4.25 bits/weight; on a small 3B model the quality loss is proportionally larger (weaker instruction adherence, more reasoning slips, worse calibration) |
| **Size**    | \~6–7 GB                                 | \~2–2.5 GB                                                                                                                                                |
| **Speed**   | Slower (weights are 4× larger to stream) | Typically 2–3× more tokens/sec                                                                                                                            |

**When to use which:** Q4\_0 for edge hardware, limited RAM/VRAM, and fast casual chat. FB16 when fidelity matters — benchmarking, structured output (JSON, code, tool calls), agentic pipelines where small errors cascade, or generating reference data.

## Disclaimer

- This build was **independently created**. The model provider is **not acting on behalf of Mistral AI** and has **no affiliation with Mistral AI whatsoever**.
- Model weights are subject to Mistral AI's license — review it before commercial use.

## License

Model weights: Mistral AI. This build and README are provided as-is, without warranty.

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