DeepReinforce's Ornith 1.0 comes in two sizes — a dense 9B and a 35B MoE (3B active), both 256K context and MIT licensed. The 9B fits comfortably on mid-range consumer GPUs at Q4.
Ornith is the open-weight model family from DeepReinforce, first released in 2026. The lineup is refreshingly simple: two sizes under the Ornith 1.0 series — a dense 9B for everyday use and a 35B MoE that only activates 3B parameters per token. Both ship with a 256K context window and a permissive MIT license, which makes them easy to self-host, fine-tune, and build products on.
The family has been gaining traction in the local LLM community as a "long-context on consumer hardware" play: the 9B runs comfortably on a mid-range GPU, while the MoE variant delivers a much larger total parameter count — with the VRAM footprint to match — at dense-3B-class speed. Below is the size ladder with Q4 VRAM math so you can pick the right card before you download anything.
Size ladder
- Ornith-1.0-9B (9B dense) — Q4 weighs about 4.95 GB; at the default 8K context it needs roughly 7.75 GB of VRAM, so any 12GB GPU (RTX 3060 12GB, RTX 4070 class) has room to spare. A 16GB card leaves generous headroom for longer contexts.
- Ornith-1.0-35B (35B total / 3B active MoE, 256 experts, 8 per token) — Q4 weighs about 19.25 GB; at 8K context you need roughly 20.9 GB, which means a 24GB GPU (RTX 3090 / 4090) is the realistic minimum. Don't even try to squeeze it into 16GB at Q4.
What about FP16?
If you want unquantized quality, FP16 weights scale directly with total parameters: the 9B needs ~18 GB, so a 24GB card covers it; the 35B MoE needs ~70 GB, which pushes you into 96GB territory (dual 48GB cards or a workstation GPU). For most users, Q4 is the sweet spot for both models.
Family highlights worth knowing
- 256K context on both sizes — unusually long for this weight class. Great for document QA and codebases, but remember KV cache grows with context: our estimates assume 8K, so budget extra VRAM if you plan to actually use 64K+ windows.
- MoE speed trick: the 35B activates only 3B parameters per token, so its tok/s lands much closer to a small dense model than to a dense 35B. You pay VRAM for all 35B parameters (they all must be resident), but you don't pay full compute for them.
- MIT license — no usage restrictions to lawyer-proof around; commercial use is fine.
- GGUF builds are the community norm for this family, so llama.cpp / Ollama / LM Studio setups are straightforward.
Which one should you pick
- 12GB VRAM (RTX 3060 12GB, 4070): take the 9B at Q4. It's the no-drama option — fits with ~4GB to spare, and that headroom lets you stretch context beyond 8K.
- 16GB VRAM (RTX 4060 Ti 16GB, 4080 16GB): the 9B is effortless here, including at much longer contexts. The 35B MoE does not fit at Q4 (needs ~21GB), so this tier stays on the 9B.
- 24GB VRAM (RTX 3090 / 4090): this is where the family gets interesting — the 35B MoE at Q4 fits with a couple of GB of headroom, giving you 35B-parameter knowledge at ~3B-class speed. The 9B also runs at FP16 if you prefer.
- 48GB (RTX A6000, dual 24GB): 35B at higher quants (Q6/Q8) or with very long contexts becomes comfortable.
- 96GB and up: FP16 for the 35B MoE, or both models loaded side by side for routing setups.
Bottom line: on 12–16GB, Ornith means the 9B — a solid, long-context daily driver. The 35B MoE is the reason to own a 24GB card in 2026.
Links
2 models in the Ornith family, grouped by series. Q4 (GB) is weights-only; total VRAM adds KV cache and overhead — the fit check links compute it for a representative retail GPU at 8K context.