Icarus-1

Icarus flies as close to the sun as your hardware allows. It's the first model in the Olympus Series from Solel//ai, and an all-rounder: chat, writing, coding, agent work and reading images, from one model. We fine-tuned it towards more conversational, lifelike replies, while keeping the strengths of the Ornith model underneath.

The Olympus Series is where we push consumer hardware to its limit. These models are sized to fill a graphics card to the brim and spill the rest into system memory, and still answer at a healthy pace. Icarus can do that because it's a mixture of experts: it holds about 35 billion parameters, but only about 3 billion of them work on each token. The parts that spill over to the processor cost far less speed than they would in a dense model.

All sunshine. No cloud.

At a glance

Family	Olympus Series by Solel//ai
Role	All-rounder
Parameters	About 35 billion in total, about 3 billion active per token (mixture of experts)
Fine-tuned from	Ornith-1.5-35B-A3B (MIT), itself built on Qwen3.5
Tuned for	More conversational, lifelike replies, with Ornith's strengths kept
Input	Text and images
Output	Text
Context window	Up to 262,144 tokens (in practice, set by your memory)
Format	GGUF in three sizes, with a Q8_0 vision projector
Licence	MIT, open weights

Files

Pick one size. Each one is the whole model; the bigger ones keep more of its quality.

File	Size	Graphics card it's sized for	Graphics and system memory together
Icarus-1.Q3_K_M.gguf	16.76 GB	16 GB	24 GB
Icarus-1.Q4_K_M.gguf	21.17 GB	20 GB	28 GB
Icarus-1.Q5_K_M.gguf	24.73 GB	24 GB	32 GB
Icarus-1.mmproj-Q8_0.gguf	0.61 GB

The projector is the vision part, needed for images. A smaller card than a size is made for still runs it, with more of the model in system memory and fewer tokens a second, as long as the two together are enough.

Sizes are decimal gigabytes. Long chats need extra memory on top for the context.

Run it

In Solel//os

Open Settings › Models. Icarus shows up under the Olympus Series on its own; pick a size, download it and choose Use. Solel//os suggests the size that suits your server.

With llama.cpp

llama-server -hf Solel-io/Icarus-1:Q4_K_M --fit on

-hf fetches the vision projector too. --fit on puts as much as fits on the graphics card and leaves the rest in system memory, experts first, which is how Icarus is meant to run.

Good at, and not so good at

Icarus is a strong all-rounder. It talks in a more natural, lifelike way than its base model, and keeps Ornith's real lean towards coding and agent work. It's slower to start and needs more memory than the Aether Series, and like every language model it can be confidently wrong. Check anything that matters.

Content and safety

Icarus has much lighter refusal behaviour than its base model. It's meant for adults running it on their own hardware. If you put it in front of other people, add your own filtering. You're responsible for how you use it and what you do with its output.

Licence

Icarus-1 is open weights under the MIT licence, the same licence as the Ornith-1.5 model it's fine-tuned from. Ornith-1.5 is built on Qwen3.5 by the Qwen team, released under Apache-2.0; that licence is kept with these files as LICENSE-QWEN.

Where the files came from, with checksums, is in provenance.json. The original model card is kept as UPSTREAM_README.md, and the MIT licence text as LICENSE.

The names Solel, Solel//ai, Olympus Series and Icarus are trade marks of Solel//io. The licence covers the weights, not the names. See the Solel//io trade mark policy.

About Solel//ai

Solel//ai is the AI division of Solel//io, an Australian open-source home server business. We make local models that run on hardware you own, so your conversations stay in your house.

• The Aether Series: Angel (fast), Dominion (everyday), Seraphim (flagship)
• The Olympus Series: Icarus (all-rounder)

solel.io · GitHub · © 2026 Jake Thomas Biddle, trading as SOLEL-IO (ABN 12 623 358 183)