CanItRun Logocanitrun.

Ornith 1.5 9B

Ornith 1.5 9B needs roughly 6.4 GB VRAM at Q4_K_M quantization (20.5 GB at FP16). 107 GPUs we track can run it fully in VRAM at 8k context.

107 GPUs run this natively · 2 with CPU offload

Ornith AI9B params256k contextMITCommercial use ok

Ornith 1.5 9B is a 9B parameter dense model developed by Ornith AI. Released 19 August 2026 under the MIT license, the lightest of three Ornith-1.5 sizes launched the same day (9B dense, 35B-A3B MoE, 397B MoE). Ornith AI continues-pretrains on top of Qwen3.5 (and, for the other sizes, Gemma 4) rather than training a new architecture: config.json is byte-identical to Qwen 3.5 9B's own hybrid stack, 32 layers built from eight repeats of three Gated DeltaNet layers followed by one Gated Attention layer, so only 8 of 32 layers keep a growing KV cache. Context is 262,144 tokens natively and roughly 1M with YaRN scaling. A separately quantized Ornith-1.5-9B-Mobile build targets on-device iPhone and Android inference.

To run Ornith 1.5 9B locally: Same shape as Qwen 3.5 9B, so the same VRAM math applies: a Q4_K_M GGUF runs a few GB of weights plus a KV cache that stays flat across most of the context window thanks to the hybrid attention stack, comfortable on any 8GB+ GPU at the recommended quant. GGUF builds are already available from ornith-ai, unsloth, and bartowski on Hugging Face, plus a native ollama library tag for anyone who'd rather not pull from Hugging Face directly.

Ornith AI's own launch comparison reports GPQA Diamond 86.4 and SWE-bench Verified 70.6, and describes the 9B as matching or exceeding much larger models on agentic coding, including Gemma 4 31B and Qwen 3.6 35B-A3B (Alibaba's MoE variant, not the dense Qwen 3.6 35B already tracked on this site), despite its size. These are Ornith AI's own figures from its launch blog and model card, measured across five averaged runs per its stated methodology; independent third-party reproduction was still thin in the first day after release.

VRAM at each quantization

Calculated at 8k context. Because only part of this model's stack keeps a growing KV cache, longer sessions cost far less extra VRAM than a full-attention model of the same size.

QuantWeightsKV cacheTotal
FP3236.0 GB0.27 GB40.6 GB
BF1618.0 GB0.27 GB20.5 GB
FP1618.0 GB0.27 GB20.5 GB
Q8_09.6 GB0.27 GB11.0 GB
Q6_K7.4 GB0.27 GB8.6 GB
Q5_K_M6.4 GB0.27 GB7.5 GB
Q4_K_Mrec5.5 GB0.27 GB6.4 GB
Q3_K_M4.3 GB0.27 GB5.2 GB
Q2_K3.4 GB0.27 GB4.1 GB
NVFP4cuda4.5 GB0.27 GB5.3 GB

KV cache is calculated at 8k context (FP16). Note that NVFP4 only runs on CUDA GPUs. Turn on TurboQuant in the calculator above for lower KV cache estimates.

Benchmarks

GPQA Diamond
86.4
SWE-bench Verified
70.6
SWE-bench Pro
47.5
Terminal-Bench 2.1
46.2

GPUs that run Ornith 1.5 9B natively (107)

Show 102 more
Plus 2 GPUs that run it with CPU offload (slower)

Notes

The lightest member of the Ornith-1.5 family, launched 19 August 2026 alongside the 35B-A3B and 397B models. Trained with Ornith AI's self-improvement loop: the model itself proposes new coding tasks, writes a task-specific scaffold for each one, and generates the solution rollouts used for reinforcement learning, rather than training against a fixed set of human-curated tasks. Ornith AI's own comparison has the 9B matching or beating much larger models on agentic coding, including Gemma 4 31B and Qwen 3.6 35B-A3B (Alibaba's MoE variant, not the dense Qwen 3.6 35B already tracked on this site). Ships a separately quantized Ornith-1.5-9B-Mobile build for on-device iPhone and Android inference. MIT licensed with no usage restrictions. GGUF quants are available from ornith-ai, unsloth, and bartowski on Hugging Face; a native ollama library tag (ornith-1.5:9b) exists alongside the raw hf.co pull.

Hugging Face ↗Ollama ↗Released 2026-08-19

How to run Ornith 1.5 9B locally

816244880160320

Q4_K_M needs 6.4 GB: fits a single high-end consumer GPU (24 GB).

llama.cpp (Ornith AI's coding-task sampling)

./llama-server \
  -hf ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M \
  -c 262144 -ngl 99 \
  --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.0

Ollama

ollama run ornith-1.5:9b

llama.cpp

# General-purpose tasks use a hotter sampling profile
./llama-server \
  -hf ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M \
  -c 262144 -ngl 99 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --presence-penalty 1.5

# Vision needs the projector file alongside the weights
./llama-server -hf ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M \
  --mmproj mmproj-Ornith-1.5-9B-BF16.gguf

LM Studio: Search for 'Ornith 1.5 9B'; if LM Studio hasn't indexed it yet, point 'Load from folder' at a downloaded Q4_K_M file from ornith-ai's own GGUF repo or bartowski's mirror. Load the separate mmproj file alongside it for image input.

Why this quantization? Q4_K_M puts the dense 9B weights at roughly 5.5 GB, small enough that even a full 262,144-token context stays well under an 8GB card's usable ceiling for most working context lengths, thanks to the hybrid stack keeping only 8 of 32 layers' KV cache growing with context. There's real headroom to size up: Q6_K or Q8_0 both still fit comfortably on a 12GB+ card if quality matters more than squeezing onto the smallest possible GPU.

Who is Ornith 1.5 9B for?

Anyone with an 8GB+ GPU (or the quantized Ornith-1.5-9B-Mobile build on a modern phone) who wants a coding-agent model that punches above its size class, per Ornith AI's own comparison against Gemma 4 31B and Qwen 3.6 35B-A3B (Alibaba's MoE variant, not the dense Qwen 3.6 35B already tracked on this site), without needing a 24GB+ card.

Best for

  • Local coding agents on modest hardware, laptops, budget GPUs, or edge devices via the Mobile build
  • Terminal- and tool-use-driven agent workflows, what Ornith AI's self-improvement training loop specifically targets
  • Long-context work where the hybrid attention stack keeps KV cache growth flat
  • A lightweight draft model for speculative decoding alongside the larger Ornith-1.5 sizes

Not ideal for

  • General knowledge or chat quality on par with larger dense models outside the coding-agent domain this release was tuned for
  • Anyone who needs benchmark claims independently verified before adopting a model, this released within days of when this page was written
  • Workloads needing a verified thinking/non-thinking toggle, Ornith AI's own docs describe sampling profiles for general vs. coding tasks rather than a chat-template thinking flag

Explore

Browse all guides →Browse all models →Browse all GPUs →

Frequently asked questions

What are the VRAM requirements for Ornith 1.5 9B?
Ornith 1.5 9B requires approximately 6.4 GB of VRAM at Q4_K_M quantization, 11.0 GB at Q8, and 20.5 GB at FP16. These numbers assume 8k context window; its hybrid attention stack caches far fewer than all layers, so VRAM grows much slower than linearly with context.
How many parameters does Ornith 1.5 9B have?
Ornith 1.5 9B has 9 billion parameters.
Can Ornith 1.5 9B run on a 16 GB GPU?
Yes. Ornith 1.5 9B needs 6.4 GB at Q4_K_M, which fits in a 16 GB GPU like the RTX 4080 or RTX 5070 Ti.
What is the smallest quantization for Ornith 1.5 9B that fits in 24 GB of VRAM?
At BF16, Ornith 1.5 9B needs 20.5 GB, the highest-quality quantization that fits in 24 GB of VRAM.
What GPU do I need to run Ornith 1.5 9B locally?
A 16 GB GPU is enough. At Q4_K_M, Ornith 1.5 9B needs 6.4 GB VRAM. Good options: RTX 4080 (16 GB), RTX 5070 Ti (16 GB).