AirgapAI Model Downloads – 16 Models Available

Q-3.5

Qwen 3.5

A capable all-rounder for everyday work, coding and images

Chat

INTELLIGENCE

Advanced

SPEED

Fast

SIZE

Small–Large

A well-rounded assistant for everyday writing, research and analysis, able to work through multi-step tasks in order and use tools along the way. It reads and writes code confidently, and can take in a screenshot or photo alongside your question. A dependable first choice for most day-to-day work.

Thinking Vision Function calling Structured output

Requires AirgapAI v2.2 or later. Make sure your app is updated before downloading these models.

Q4_K_MGPU/CPU2.0 GB INT4GPU/NPU2.1 GB

Q4_K_MGPU/CPU3.5 GB INT4GPU/NPU3.5 GB

Q4_K_MGPU6.8 GB Q4_K_MGPU/CPU5.6 GB INT4GPU6.1 GB

27B

Q4_K_MGPU18 GB

35B-A3B

Q4_K_MGPU24 GB

Q4Apple Silicon2.0 GB

Q4Apple Silicon3.5 GB

Q4Apple Silicon6.8 GB

27B

Q4Apple Silicon18 GB

35B-A3B

Q4Apple Silicon24 GB

Reads images and screenshots alongside text
4B is the easy default; 9B when answer quality matters most
Choose a size, then the download row under it
A second 9B Windows row offers uncensored weights (fewer refusals); use only where that trade-off is appropriate.
Learn more (4B)
Learn more (9B)

Q-3.8

Qwen 3.8 Distill

A quick and capable all-rounder for everyday work

Chat

INTELLIGENCE

Advanced

SPEED

Very Fast

SIZE

Small–Medium

A compact assistant that reasons in visible steps before it answers, so you can follow how it reached a conclusion. It loads quickly and responds quickly, which suits everyday questions, drafting and code on modest hardware.

Thinking Function calling Structured output

Requires AirgapAI v2.2 or later. Make sure your app is updated before downloading these models.

Q4_K_MGPU/CPU1.3 GB

Q4_K_MGPU/CPU2.8 GB

Q4_K_MGPU/CPU5.8 GB

Q4Apple Silicon1.3 GB

Q4Apple Silicon2.8 GB

Q4Apple Silicon5.8 GB

9B when answer quality matters most; 4B for quick, smooth responses; 2B for the smallest machines
Text-only build: no image input
Apache-2.0 licensed
Choose a size, then the download row under it
Learn more (9B)
Learn more (4B)

GEMMA

Gemma 4

Reads pictures and scanned pages, and writes naturally

Chat

INTELLIGENCE

Advanced

SPEED

Fast

SIZE

Small–Large

Works with images as fluently as it works with text: show it a photo, a screenshot or a scanned page and it will describe what is there or pull out specific details. It writes fluent, natural prose, which suits longer documents, emails and creative work.

Thinking Vision Function calling Structured output

Requires AirgapAI v2.2 or later. Make sure your app is updated before downloading these models.

E2B

Q4_K_MGPU/CPU4.1 GB INT4GPU/NPU4.4 GB

E4B

Q4_K_MGPU/CPU6.0 GB INT4GPU/NPU6.5 GB

12B

Q4_K_MGPU7.3 GB

26B-A4B

Q4_K_MGPU18 GB

31B

Q4_K_MGPU20 GB

E2B

Q4Apple Silicon4.1 GB

E4B

Q4Apple Silicon6.0 GB

12B

Q4Apple Silicon7.3 GB

26B-A4B

Q4Apple Silicon18 GB

31B

Q4Apple Silicon20 GB

Reads images, screenshots and scanned pages alongside text
4-bit quantisation across all variants
E4B runs on most modern GPUs / iGPUs; larger builds want a 2022+ discrete GPU or Apple Silicon
Choose a size, then the download row under it
Model details

L-3.2

Llama 3.2

A very light assistant for modest hardware

Legacy Chat

INTELLIGENCE

Capable

SPEED

Very Fast

SIZE

Small

A compact assistant built to run on machines with little memory to spare. It answers quickly and suits short questions, quick rewrites and simple summaries.

INT4NPU

Q4_K_MGPU/CPU2.0 GB INT4GPU INT4NPU

Q4Apple Silicon2.0 GB

Includes a longer-context 3B build for working through long documents
4-bit quantisation across all variants
1B runs on older GPUs / iGPUs; 3B benefits from 2025 Intel iGPU or 2022+ discrete GPU
Model details (1B)
Model details (3B)

L-8B

Llama 3.1 8B

A steady general assistant for conversation and writing

Legacy Chat

INTELLIGENCE

Strong

SPEED

Fast

SIZE

Medium

A general-purpose assistant for everyday conversation, drafting and summarising. It stays coherent across longer pieces of text, which suits routine writing and question-answering.

Q4_K_MGPU4.9 GB INT4GPU INT4NPU

Q4Apple Silicon4.9 GB

Built for Intel hardware: choose the GPU build for Arc / integrated graphics, or NPU on Core Ultra laptops
INT4 quantisation
Requires Lunar Lake or later for best results
Model details

SAUL

Saul 7B Instruct

A legal specialist for contracts, clauses and case language

Legal specialist Chat

INTELLIGENCE

Strong

SPEED

Fast

SIZE

Medium

Trained on legal writing, so it reads contracts, clauses and obligations with an eye for why a particular phrase matters. Suited to contract review, case summaries and legal questions. It is not a substitute for advice from a qualified lawyer.

Q4_K_MGPU4.4 GB

Q4Apple Silicon4.4 GB

Built on Mistral 7B, further trained on a large body of legal text
5 GB+ vRAM
Runs via llama.cpp runtime (GPU/CPU hybrid supported)
Specialised for legal reasoning and drafting
Model details

AFM

Arcee AFM 4.5B

A compact general assistant tuned for Intel machines

Chat

SPEED

Fast

SIZE

Small

A small general-purpose assistant that runs comfortably on Intel laptops and desktops without a separate graphics card. It handles everyday questions, short drafts and summaries.

4.5B

INT8GPU

5 GB+ vRAM
Optimised for 2024+ Intel Arc / iGPU
Model details

GR4.2

IBM Granite 4.2

Document answers with visible reasoning and tool use

Chat

INTELLIGENCE

Strong

SPEED

Very Fast

SIZE

Small–Medium

Answers questions from documents you supply, reasoning in visible steps so you can follow its working. It can call on a tool part-way through an answer when something needs looking up first.

Thinking Function calling Structured output

Q4_K_MGPU/CPU2.2 GB

Q4_K_MGPU/CPU5.3 GB

Q4Apple Silicon2.2 GB

Q4Apple Silicon5.3 GB

8B is the quality pick; 3B for fast responses on more machines
Multilingual: tested in English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese
Apache-2.0 licensed
Choose a size, then the download row under it
Learn more (8B)
Learn more (3B)

IBM-G

IBM Granite 4.1

Answers from your own documents, with sources you can check

Chat

INTELLIGENCE

Strong

SPEED

Fast

SIZE

Medium

Built to answer questions from documents you supply — reports, contracts, policies, handbooks. It stays close to what those sources actually say, so its answers are easy to trace back and verify.

Q4_K_MGPU5.3 GB

Q4Apple Silicon5.3 GB

Single 8B size
Text only: no image input
Tuned for question-answering over documents you supply
Model details

BLK

By Iternal Technologies

Blockify

Turns your documents into clean, searchable pieces

Document Processing

SPEED

Very Fast

SIZE

Small–Medium

MODALITY

Text → IdeaBlocks

Blockify prepares documents for search: it reads them and breaks them into small, self-contained pieces of meaning, so later questions are matched against clean material rather than raw pages. Iternal built and trained it for this job. Review the output before relying on it for anything important.

INT4GPU

IQ4_XSGPU/CPU1.2 GB

INT4GPU INT4NPU INT4CPU INT8GPU INT8NPU

26B-A4B

IQ2_MGPU/CPU5.7 GB

IQ4_XSApple Silicon1.2 GB

26B-A4B

IQ2_MApple Silicon5.7 GB

Processes documents into optimised IdeaBlocks
GGUF builds run on any GPU with CPU fallback: the 2B needs about 2.6 GB of VRAM, the Gemma 4 26B-A4B build about 7 GB
OpenVINO builds are optimised for Intel: 1B on ≥ 2 GB vRAM; 8B on ≥ 5 GB
Read more about Blockify here

Jina Embeddings V2

Powers search that understands meaning, not just words

Vector Search

SPEED

Instant

MEMORY

2 GB+

Turns text into a form that can be compared by meaning, so document search finds the right passage even when your wording differs from the source. It runs underneath search and retrieval, quickly and with a light footprint.

CPU691 MB

Uses CPU and normal Memory
Model details

NLLB

NLLB Text Translation

Translates text across more than 200 languages, offline

Translation

SPEED

Very Fast

MODALITY

Text → Text

Translates text between over 200 languages, including many that mainstream tools do not cover. It runs entirely on your machine, so nothing you translate leaves the device.

CPU

Supports 200+ languages
Uses CPU and normal Memory
Model details

IBM Granite Speech 4.1

High-accuracy transcripts in English and five other languages

Audio to Text

MEMORY

8 GB+

MODALITY

Audio → Text

IBM's speech model for accurate, punctuated transcripts in English, French, German, Spanish, Portuguese and Japanese, and it can translate speech as well. This 4-bit build runs on llama.cpp and is fastest on a capable graphics card.

Q4_K_MGPU/CPU2.3 GB

Q4_K_MApple Silicon2.3 GB

4-bit (Q4_K_M) weights with the matching audio projector (mmproj) included
Needs llama.cpp build b9045 or newer
Runs on the GPU where available, otherwise on the CPU
Install in AirgapAI Transcribe by importing the downloaded zip
Model details

Nemotron 3.5 Multilingual (Foundry Local)

Transcribe's built-in speech model, run through Microsoft Foundry Local

Audio to Text

MEMORY

4 GB+

MODALITY

Audio → Text

NVIDIA's Nemotron speech model as Microsoft packages it for Foundry Local. It is the same model AirgapAI Transcribe has built in, run a different way, for computers where the built-in model will not run. It transcribes English and 10 other languages on the processor, with times by sentence rather than by word.

Requires a version of AirgapAI Transcribe that includes the Foundry Local runtime.

0.6B

CPU793 MB

NVIDIA Nemotron 3.5 ASR Streaming 0.6B in ONNX format, from the Foundry Local catalog (nemotron-3.5-asr-streaming-0.6b-generic-cpu:3)
Languages: English, Chinese, Dutch, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish
Runs on the CPU on Windows
Download it in AirgapAI Transcribe from Settings > Speech model > More speech models
Licenses as listed in the download's NOTICES file: NVIDIA Open Model License for the weights, MIT License for the Foundry Local packaging and Silero VAD
Model details
NVIDIA Open Model License

Speaker Labels

Marks who spoke when in a transcript

Speaker Detection

MODALITY

Audio → Speaker labels

Attributes each passage of an AirgapAI Transcribe transcript to a speaker, so meetings and interviews read as a conversation. It is small and runs on the CPU. The Transcribe installer already includes it; download it here only to restore the files if they are missing.

CPU26 MB

Contains the Transcribe installer bundle: diarization-model.bundle.tar and its manifest
Segmentation: pyannote segmentation 3.0 (INT8, MIT License)
Speaker embedding: WeSpeaker, Creative Commons Attribution 4.0
Segmentation model details
WeSpeaker project

WL3

Whisper Large V3

Transcribes speech in 99 languages, offline

Audio to Text

MEMORY

4 GB+

MODALITY

Audio → Text

Turns recordings into text in 99 languages and can translate speech into English, with word-level timings in the transcript. This 4-bit build is under a third of the full model's size and runs on a graphics card or the CPU.

Large

Q4_0GPU/CPU889 MB

Large

Q4_0Apple Silicon889 MB

OpenAI Whisper large-v3 in whisper.cpp format, 4-bit (Q4_0)
Runs on the GPU where available, otherwise on the CPU
Install in AirgapAI Transcribe by importing the downloaded zip
Model details

AirgapAI Local Model Downloads

Large Language Models

Qwen 3.5

Qwen 3.8 Distill

Gemma 4

Llama 3.2

Llama 3.1 8B

Saul 7B Instruct

Arcee AFM 4.5B

IBM Granite 4.2

IBM Granite 4.1

Document Processing

Blockify

Embeddings & Vector Search

Jina Embeddings V2

Translation

NLLB Text Translation

Speech & Audio

IBM Granite Speech 4.1

Nemotron 3.5 Multilingual (Foundry Local)

Speaker Labels

Whisper Large V3