Loading...
 
Skip to main content
OpenSource Solutions model documentation

Open source AI models we run and test


These are the models installed on our servers for the private AI assistants we build. Each one is tested on real business questions before we recommend it. None of them sends your data to a third party.


Test them on your questionsPrivate AI agents
Google

gemma4:12b

Completion Thinking Tools Vision

Google's Gemma 4 12B: a unified multimodal model without separate encoders, reading text, images and audio, sized for a laptop. Sits between the mobile E4B and the larger 26B and 31B models.

11.9B
Parameters
7.6 GB
Memory
49
Tokens/s
License: Apache 2.0
Google

gemma4:12b-it-q8_0

Completion Thinking Tools Vision

Google's Gemma 4 12B in 8-bit quantization (Q8_0): the same model as gemma4:12b with less rounding, so closer to the original, at 12.8 GB instead of 7.6 GB. Shows on the bench whether the extra memory buys better ans...

11.9B
Parameters
12.8 GB
Memory
-
Tokens/s
License: Apache 2.0
Google

gemma4:12b-it-qat

Completion Thinking Tools Vision

Google's Gemma 4 12B in its quantization-aware training (QAT) build: trained to be stored in 4 bits (Q4_0), so it keeps quality close to the full-precision model at about 7 GB. Same model as gemma4:12b otherwise: tex...

11.9B
Parameters
7.2 GB
Memory
-
Tokens/s
License: Apache 2.0
Google

gemma4:e4b

Completion Thinking Tools Vision

Gemma 4 E4B (effective 4 billion parameters): Google's model for phones and small devices, reading text, images, video and audio. Fast, but limited on detailed answers.

8.0B
Parameters
9.6 GB
Memory
97
Tokens/s
License: Apache 2.0
OpenAI

gpt-oss:20b

Completion Thinking Tools

OpenAI's open-weight reasoning model: 21 billion parameters, of which 3.6 billion are active per token (mixture of experts). Built to run on a single 16 GB graphics card, with tool use and adjustable reasoning effort...

20.9B
Parameters
13.8 GB
Memory
95
Tokens/s
License: Apache 2.0
CogEvol, quantization by bartowski

hf.co/bartowski/CogEvol_CogEvol-4B-GGUF:Q4_K_M

Completion Thinking Tools Vision

Fine-tune of Qwen3.5-4B specialised in generating teaching material: it turns a course brief into slide pages (JSON) or interactive HTML pages. Not a general assistant, which explains its low scores on general questi...

4.21B
Parameters
3.6 GB
Memory
113
Tokens/s
License: Apache 2.0
ukisai (based on Alibaba Qwen3.8-27B)

hf.co/ukisai/Swift-1.5-Qwen3.8-27B-GGUF:Q3_K_M

Completion Thinking Tools Vision

Fine-tune of Qwen3.8-27B by ukisai, trained to reason with far fewer thinking tokens (about 58% fewer) for the same score, so it answers faster on long agent and coding tasks.

27.3B
Parameters
14.3 GB
Memory
27
Tokens/s
License: Swift Open License v1.0: free up to USD 1M annual revenue...
Alibaba (Qwen)

hf.co/unsloth/Qwen3.6-27B-GGUF:Q3_K_M

Completion Thinking Tools Vision

Alibaba's 27-billion-parameter dense model, reads text and images, with optional reasoning. Strong at code and agent tasks for its size. Here in a 3-bit quantization (Q3_K_M) so that it fits a 16 GB card.

26.9B
Parameters
14.5 GB
Memory
27
Tokens/s
License: Apache 2.0
Alibaba (Qwen), quantization by Unsloth

hf.co/unsloth/Qwen3.8-27B-GGUF:UD-Q3_K_XL

Completion Thinking Tools Vision

Successor of Qwen3.6-27B with the same architecture and new post-training, aimed at agent work. Reasons more by default, which makes it slower. Here in Unsloth's 3-bit dynamic quantization (UD-Q3_K_XL).

27.3B
Parameters
14.1 GB
Memory
29
Tokens/s
License: Apache 2.0
Liquid AI

lfm2.5:8b

Completion Thinking Tools

Liquid AI's device-oriented mixture-of-experts model: 8.3 billion parameters, about 1.5 billion active, hybrid convolution and attention architecture, 128K context. Very fast, reasons before answering, good at tool c...

8.5B
Parameters
5.2 GB
Memory
276
Tokens/s
License: LFM Open License v1.0: free under USD 10M annual revenue,...
Mistral AI

ministral-3:14b

Completion Tools Vision

Mistral AI's 14-billion-parameter model of the Ministral 3 family (December 2025), here the instruct version: text and images, multilingual, with tool use, no reasoning mode. Built for local and edge deployment; fits...

13.9B
Parameters
9.1 GB
Memory
-
Tokens/s
License: Apache 2.0
Ornith AI

ornith-1.5:9b

Completion Thinking Tools Vision

Ornith AI's 9-billion-parameter model, the smallest of the Ornith-1.5 family, tuned for terminal-based coding agents and tool calling; reads text, images and video, 262K-token context. The family is built on Qwen3.5 ...

9.0B
Parameters
6.6 GB
Memory
-
Tokens/s
License: Not stated (checked 16 October 2026)
Alibaba (Qwen)

qwen3-vl:8b

Completion Thinking Tools Vision

Vision-language model of Alibaba's Qwen3-VL family, 8.8 billion parameters: reads images and documents, spatial understanding, long context. Older generation than Qwen3.5.

8.8B
Parameters
6.1 GB
Memory
81
Tokens/s
License: Apache 2.0
Alibaba (Qwen)

qwen3.5:9b

Completion Thinking Tools Vision

Small model of Alibaba's Qwen3.5 family: about 9 billion parameters, reads text and images, reasons before answering, 262K-token context. Designed to run on a laptop.

9.7B
Parameters
6.6 GB
Memory
71
Tokens/s
License: Apache 2.0

Which model for your assistant?


A model alone is not an assistant. OpenSource Solutions builds the harness around it: the connection to your documents and tools, the rules it follows, the hosting on your servers or ours, and the measurement on your own questions before you choose. Your data stays on infrastructure you control.


Start with an AI Readiness AuditPrivate AI agents