Kasoku

Models

Pick your tier to see what fits. Each model has the exact Ollama command or LM Studio file for your memory.

19/23run on a Solid card

Showing a typical Solid card. Test yours

Best pick for this card

Qwen3.5 4B

Chat, 4B, Alibaba Qwen

First model to install on a 4 to 6 GB card: fits with room for long chats and images.

Operating system
Download
3.3 GB
Memory
4 GB
Speed
~34 tok/s estimate

Ollama

Free, runs from the command line.

  1. 1. Download Ollama for Windows and run the installer.
  2. 2. Open PowerShell and run:
    ollama run qwen3.5:4b
    Q4_K_M, 3.3 GB download. The first run downloads it, then the chat starts.

LM Studio

Free desktop app with a chat window.

  1. 1. Download LM Studio and install it.
  2. 2. Open the model search and look for:
    Qwen3.5-4B
  3. 3. In lmstudio-community/Qwen3.5-4B-GGUF, download this file:
    Qwen3.5-4B-Q4_K_M.gguf
    Q4_K_M, 2.7 GB download
  4. 4. Load it in the Chat tab and start typing.

We do not download or install anything. The links go to the official apps.

Runs great

15Fast enough to use every day.
  • Gemma 3 1BChat, Google Gemma

    Tiny chat model for quick answers, rewrites and short summaries on almost any machine.

    1B F16

    2.0 GB download

    3 GB memory

    ~75 tok/sestimate

  • Llama 3.2 3BChat, Meta Llama

    Small general chat model that handles summaries, rewriting and simple questions well.

    3B Q8_0

    3.4 GB download

    4.5 GB memory

    ~45 tok/sestimate

  • Phi-4-miniChat, Microsoft

    Compact model that is good at math, logic and following instructions for its size.

    3.8B Q8_0

    4.1 GB download

    5.5 GB memory

    ~38 tok/sestimate

  • Qwen3.5 4BChat, Alibaba Qwen

    Current small Qwen model with a thinking mode, many languages and basic image input.

    4B Q4_K_M

    3.3 GB download

    4 GB memory

    ~34 tok/sestimate

  • Moondream 2Vision, Moondream

    Tiny vision model that captions photos and answers simple questions about them.

    1.9B F16

    3.8 GB download

    5 GB memory

    ~55 tok/sestimate

  • Qwen3-VL 4BVision, Alibaba Qwen

    Vision model that describes images, reads text in screenshots and answers questions about charts.

    4B Q4_K_M

    3.3 GB download

    4.5 GB memory

    ~32 tok/sestimate

  • Stable Diffusion 1.5Image, Stability AI / Runway

    Classic 512px image generator with a huge library of community styles and fine-tunes.

    0.9B fp16

    2 GB download

    3.5 GB memory

    ~6 s/imageestimate

  • Whisper tinySpeech, OpenAI

    Fastest speech to text model; fine for clear English, weaker on accents and noise.

    39M fp32

    0.1 GB download

    0.5 GB memory

    ~60x realtimeestimate

  • Kokoro 82MSpeech, hexgrad

    Small text to speech model with natural sounding voices in English and a few other languages.

    82M fp32

    0.3 GB download

    0.5 GB memory

    ~20x realtimeestimate

  • Whisper smallSpeech, OpenAI

    Balanced speech to text model with good accuracy in many languages.

    244M fp16

    0.5 GB download

    1.5 GB memory

    ~20x realtimeestimate

  • Near large-v3 accuracy for transcription at several times the speed.

    809M fp16

    1.6 GB download

    3 GB memory

    ~12x realtimeestimate

  • Swin2SR x4Tools, University of Wurzburg

    Transformer upscaler that enlarges small images four times with sharp detail.

    14M fp32

    0.1 GB download

    2 GB memory

    ~5 s/runestimate

  • Real-ESRGAN x4plusTools, Xintao Wang (Tencent ARC)

    Upscales photos four times and cleans up blur and compression noise.

    17M fp32

    0.1 GB download

    1.5 GB memory

    ~3 s/runestimate

  • RMBG-1.4Tools, BRIA AI

    Removes the background from photos of people, products and objects.

    44M fp32

    0.2 GB download

    0.5 GB memory

    ~0.5 s/runestimate

  • nomic-embed-text v1.5Tools, Nomic AI

    Turns text into search vectors so you can find the right passage in your PDFs and notes.

    137M F16

    0.3 GB download

    0.6 GB memory

    ~3 s/runestimate

Runs, but slow

4Works. Expect to wait.
  • Qwen2.5-Coder 7BChat, Alibaba Qwen

    Coding model that writes, explains and fixes code in most popular languages.

    7B Q4_K_M

    4.7 GB download

    6 GB memory

    ~21 tok/sestimate

  • Llama 3.1 8BChat, Meta Llama

    Widely used 8B chat model with a 128k context window and broad tool support.

    8B Q4_K_M

    4.9 GB download

    6 GB memory

    ~20 tok/sestimate

  • Reasoning model distilled from DeepSeek-R1 that thinks step by step before it answers.

    8B Q4_K_M

    5.2 GB download

    6.5 GB memory

    ~19 tok/sestimate

  • Gemma 4 E4BChat, Google Gemma

    Google's on-device model that runs with the compute of a 4B model and understands images and audio.

    E4B (8B total) Q4_K_M

    6.6 GB download

    7.5 GB memory

    ~26 tok/sestimate

Not for this card yet

4Needs more memory or a faster chip.
  • Qwen3.5 9BChat, Alibaba Qwen

    Strongest small all-rounder in this list for writing, coding, translation and image questions.

    9B Q4_K_M

    6.5 GB download

    8 GB memory

    ~16 tok/sestimate

  • gpt-oss 20BChat, OpenAI

    OpenAI's open-weight reasoning model; a mixture of experts, so it runs faster than its size suggests.

    21B (3.6B active) MXFP4

    13.8 GB download

    16 GB memory

    From Beast

  • SDXL-TurboImage, Stability AI

    Fast SDXL variant that makes a 512px image in one to four steps.

    3.5B fp16

    6.9 GB download

    8 GB memory

    ~4 s/imageestimate

  • FLUX.1 schnellImage, Black Forest Labs

    High quality 1024px images with readable text, in four steps.

    12B Q4_K_S (GGUF)

    12 GB download

    12 GB memory

    From Strong