Skip to content
AI Compute Radar
IndexCollected 5 h ago

What should I use? · A handful, not a thousand

Say what you want
to do. Get three.

Pick the job, confirm your machine, and you get at most three open models we would take for it, each with the reason, the file and whether it runs on your card. No scores on this page. The numbers behind the choice sit behind one link.

◇ Estimated— Estimated: computed from our curated model and hardware catalog — not a live reading.

1 What do you want to do?

2 Your machine

Nothing is sent. The card lives in your browser.

3 Our three

Order by

Write & chat · on your RTX 4090 · lists checked 2026-11-09

Qwen3.8 27B. Runs at 8K, barely fits at 32K.

5 of 5 on our write & chat list run on your RTX 4090. Order: whether it runs first, then most popular.

  1. 1 of 5more popular ▲

    Qwen3.8 27BAlibaba Qwen · 27.8B · Q4_K_M · released Aug 2026

    Runs

    Alibaba's dense 27B of the Qwen3.8 line: the most downloaded model of its size on the Hub in the last 30 days, and a 262K window.

    File
    15.3 GB
    At 8K
    17.1 GB · runs
    At 32K
    19.4 GB · barely fits
    Tools
    llama.cpp, LM Studio, Ollama
    Show the numbers
    30-day downloads
    7,232,703
    Popularity
    more popular ▲
    On the list since
    2026-11-03
    Start it in one line
    • llama.cppllama-server -hf unsloth/Qwen3.8-27B-GGUF:Q4_K_M
    • Ollamaollama run hf.co/unsloth/Qwen3.8-27B-GGUF:Q4_K_M

    Both lines name the repo and quantisation of the measured file; the tool fetches it from the Hub.

  2. 2 of 5about the same ◆

    Gemma 4 E4BGoogle · 8B · Q4_K_M · released Mar 2026

    Runs easily

    Google's small Gemma 4 for 8 GB cards: eight billion on disk, four active in memory, this week's pick of the week.

    File
    4.6 GB
    At 8K
    5.8 GB · runs easily
    At 32K
    7.2 GB · runs easily
    Tools
    llama.cpp, LM Studio, Ollama
    Show the numbers
    30-day downloads
    4,277,843
    Popularity
    about the same ◆
    On the list since
    2026-11-03
    Start it in one line
    • llama.cppllama-server -hf unsloth/gemma-4-E4B-it-GGUF:Q4_K_M
    • Ollamaollama run hf.co/unsloth/gemma-4-E4B-it-GGUF:Q4_K_M

    Both lines name the repo and quantisation of the measured file; the tool fetches it from the Hub.

  3. 3 of 5about the same ◆

    Ornith 1.5 9BOrnith AI · 9.7B · Q4_K_M · released Aug 2026

    Runs easily

    Ornith's dense 9.7B from August 2026, a 5 GB file with a 262K window; a pick of the week in October.

    File
    5.4 GB
    At 8K
    6.7 GB · runs easily
    At 32K
    8.2 GB · runs easily
    Tools
    llama.cpp, LM Studio, Ollama
    Show the numbers
    30-day downloads
    523,602
    Popularity
    about the same ◆
    On the list since
    2026-11-03
    Start it in one line
    • llama.cppllama-server -hf ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M
    • Ollamaollama run hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M

    Both lines name the repo and quantisation of the measured file; the tool fetches it from the Hub.

The list moves

Checked every Monday with the catalog run; last check 2026-11-09. The order and the fit update twice a day from the radar.

Your card is remembered

Come back with a new job and the answer is one click away. Change the card any time.

Stored in your browser only.

Tell me when a list changes

The Monday report carries one line per list that changed, with your card.

Monday mail

How the three are chosen

Each job has a list of at most 5 open models, kept by hand with a date and a reason for every entry, and published below. A model gets on a list only if its weights are open, it runs in a common tool (llama.cpp, LM Studio, Ollama), it sits among the most downloaded of its kind on Hugging Face, and it was released within the last 12 months; a successor replaces its predecessor once it meets the rule. Nothing on a list is a quality verdict. On your machine the page orders the list: first by how the file fits your card at 8K and 32K, then by popularity direction and 30-day downloads, or by file size when you ask for quick replies or the biggest that fits. The first 3 are shown; the rest are a click away. Every Monday the catalog run proposes additions that meet the rule and removals after 3 weeks of fading; a person accepts or rejects each one, and the page says when the lists were last checked.

Show today's lists (checked 2026-11-09)
Write & chat
Qwen3.8 27B · Gemma 4 31B · Gemma 4 12B · Gemma 4 E4B · Ornith 1.5 9B
Code
Qwen3.8 27B · Devstral Small 2 24B · GLM-4.7-Flash · Ornith 1.5 35B-A3B · Granite 4.2 8B
Long documents
Qwen3.8 27B · Gemma 4 31B · Gemma 4 12B · Gemma 4 E4B · Ornith 1.5 9B
Train a model
Qwen3.8 27B · Devstral Small 2 24B · Granite 4.2 8B · Granite 4.2 3B · LFM2.5 1.2B
Make images
Qwen-Image 2.1 Turbo · Z-Image Turbo · MOSS-Image · Krea 2 Turbo · FLUX.2 Klein 4B
Make video
LTX-2.5 · MiniMax-H3 · HunyuanVideo 1.5

Keep reading