CanItRun Logocanitrun.

Ornith 1.5 35B-A3B (MoE)

Ornith 1.5 35B-A3B (MoE) needs roughly 24.1 GB VRAM at Q4_K_M quantization (78.6 GB at FP16). 78 GPUs we track can run it fully in VRAM at 8k context.

78 GPUs run this natively · 21 with CPU offload

Ornith AI35B params3B active (MoE)256k contextMITCommercial use ok
Q4_K_M total
24.1 GB
at 8k context
Smallest GPU
32 GB
NVIDIA RTX 5090, at Q4_K_M
KV cache, full context
5.4 GB
10 of 40 layers cache
Inputs
Text · Image · Video
MIT, released August 2026
Active parameters
~3B
of 35B total (MoE)
Active experts
8 routed + 1 shared
of 256 total

Ornith 1.5 35B-A3B (MoE) is a Mixture of Experts (MoE) model with 35B total parameters but only 3B active per token developed by Ornith AI. Released 19 August 2026 under the MIT license alongside the 9B and 397B Ornith-1.5 sizes. Pairs MoE routing (256 experts, 8 routed plus 1 shared active, roughly 3B active parameters per token) with the same hybrid attention idea as the 9B: 40 layers built from ten repeats of three Gated DeltaNet layers followed by one Gated Attention layer, so only 10 of 40 layers keep a growing KV cache. config.json is byte-identical to Qwen 3.5 35B-A3B's own, the base checkpoint Ornith AI continues-pretrains on top of. Context is 262,144 tokens natively and roughly 1M with YaRN scaling.

To run Ornith 1.5 35B-A3B (MoE) locally: Same shape as Qwen 3.5 35B-A3B, so the VRAM story is the same: only the total weights (all 256 experts) count toward the download size, while the active-parameter count keeps decode fast, and the hybrid stack keeps KV cache growth flat across most of the context window. GGUF, MLX, FP8, and NVFP4 builds are already up from ornith-ai on Hugging Face, plus community GGUF requants from bartowski and unsloth. As a MoE model, inference speed depends on active parameters (3B) rather than total size.

Ornith AI's own launch post is explicit about the headline claim: despite activating only about 3B parameters per token, the 35B-A3B beats dense peers already tracked on this site by wide margins on agentic coding, Terminal-Bench 2.1 68.5 versus Gemma 4 31B's 43.4 and Muse Glimmer 30B's 51.7, and SWE-bench Verified 79.0 versus Gemma 4 31B's 52.0 and Muse Glimmer 30B's 76.0. It also outperforms Qwen 3.6 35B-A3B, Alibaba's similarly-sized MoE model (not the dense Qwen 3.6 35B already tracked on this site), across coding and agentic benchmarks. These are Ornith AI's own comparison figures; 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
FP32140.0 GB0.17 GB157.0 GB
BF1670.0 GB0.17 GB78.6 GB
FP1670.0 GB0.17 GB78.6 GB
Q8_037.2 GB0.17 GB41.9 GB
Q6_K28.7 GB0.17 GB32.4 GB
Q5_K_M24.9 GB0.17 GB28.1 GB
Q4_K_Mrec21.3 GB0.17 GB24.1 GB
Q3_K_M16.8 GB0.17 GB19.0 GB
Q2_K13.3 GB0.17 GB15.1 GB
NVFP4cuda17.5 GB0.17 GB19.8 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
89.2
SWE-bench Verified
79.0
SWE-bench Pro
59.6
Terminal-Bench 2.1
67.8

Beats a same-day dense peer already on this site

Ornith AI's own launch comparison names Gemma 4 31B and Muse Glimmer 30B as the dense models it benchmarks the 35B-A3B against. Muse Glimmer 30B is the closer match to verify here, both a local coding-agent release from the same August 2026 window, with scores already independently sitting in this site's own data rather than freehand-quoted from Ornith AI's table.

Terminal-Bench 2.1
Ornith 1.5 35B-A3B
67.8
Muse Glimmer 30B
51.7
SWE-bench Verified
Ornith 1.5 35B-A3B
79.0
Muse Glimmer 30B
76.0

Ornith AI, Ornith-1.5 model card and launch blog (19 August 2026), https://ornith.ai/ornith_1_5.html. The Ornith 1.5 score is Ornith AI's own vendor-reported figure; Muse Glimmer 30B's score is Meta's own self-reported figure, already tracked in this site's data. Neither has yet been independently reproduced by a third party.

Muse Glimmer 30B runs all 27.8B of its parameters on every token; Ornith 1.5 35B-A3B activates only about 3B. It still leads on both benchmarks: Terminal-Bench 2.1 67.8 vs 51.7, and SWE-bench Verified 79.0 vs 76.0. That gap is the practical argument for MoE routing in one chart, a fraction of the active compute and still ahead, not just a theoretical efficiency win.

GPUs that run Ornith 1.5 35B-A3B (MoE) natively (78)

Show 73 more
Plus 21 GPUs that run it with CPU offload (slower)

Notes

256 experts, 8 routed + 1 shared active per token (~3B active parameters total). Launched 19 August 2026 alongside the 9B and 397B models, trained with the same self-improvement loop described on the 9B entry. Per Ornith AI's own published comparison, despite activating only ~3B parameters per token it beats dense peers already on this site by wide margins on agentic coding: Terminal-Bench 2.1 68.5 vs Gemma 4 31B's 43.4 and Muse Glimmer 30B's 51.7 (Claude-Code-harness scores; see this entry's benchmarks field for the Terminus-2 numbers), and SWE-Bench Verified 79.0 vs Gemma 4 31B's 52.0 and Muse Glimmer 30B's 76.0. It also outperforms Qwen 3.6 35B-A3B, Alibaba's similarly-sized MoE model (not the dense Qwen 3.6 35B already tracked on this site), across coding and agentic benchmarks. MIT licensed. GGUF, MLX, FP8, and NVFP4 builds are available from ornith-ai on Hugging Face, plus community GGUF requants from bartowski and unsloth.

Hugging Face ↗Ollama ↗Released 2026-08-19

Compare Ornith 1.5 35B-A3B (MoE) with other models

How to run Ornith 1.5 35B-A3B (MoE) locally

816244880160320

Q4_K_M needs 24.1 GB: needs a workstation or datacenter GPU (48–80 GB).

llama.cpp (Ornith AI's recommended sampling)

./llama-server \
  -hf ornith-ai/Ornith-1.5-35B-A3B-GGUF:Q4_K_M \
  -c 262144 -ngl 99 \
  --temp 0.6 --top-p 0.95 --top-k 20

Ollama

ollama run ornith-1.5:35b

llama.cpp

# To reproduce Ornith AI's own published benchmark numbers
./llama-server \
  -hf ornith-ai/Ornith-1.5-35B-A3B-GGUF:Q4_K_M \
  -c 262144 -ngl 99 \
  --temp 1.0 --top-p 0.95 --top-k 20

# Vision needs the projector file alongside the weights
./llama-server -hf ornith-ai/Ornith-1.5-35B-A3B-GGUF:Q4_K_M \
  --mmproj mmproj-Ornith-1.5-35B-BF16.gguf

LM Studio: Search for 'Ornith 1.5 35B A3B' and take the Q4_K_M build from ornith-ai or bartowski. Because most of the total size is MoE weights rather than KV cache, raising the context slider costs far less VRAM here than on a same-sized dense model; budget for the weights floor first, then add context.

Why this quantization? Q4_K_M puts the 256-expert weights at roughly 21.3 GB, close to what a 24GB card can hold once activation overhead is added, so a 32GB card is the comfortable target, same as the Qwen 3.5 35B-A3B checkpoint it's continues-pretrained from. The hybrid attention stack keeps context cheap once that weights floor is cleared: the full 262,144-token window adds only a few GB of KV cache on top, since just 10 of 40 layers actually grow their cache with context.

Who is Ornith 1.5 35B-A3B (MoE) for?

Anyone with a 32GB GPU (or 32GB+ unified-memory Apple Silicon) who wants a coding-agent model that, per Ornith AI's own launch comparison, beats dense models roughly its own size on agentic coding despite activating only about 3B parameters per token.

Best for

  • Local coding agents that need more headroom than the 9B but don't want a full dense-35B compute cost
  • Long-document and multi-file analysis, where the hybrid attention stack keeps KV cache growth flat
  • Agentic and tool-use workflows that benefit from MoE-level throughput without a datacenter GPU
  • Image and video understanding via the vision projector

Not ideal for

  • 16-24GB GPUs, which will offload part of the 35B weight footprint even at short context
  • Anyone who needs the ~3B active-parameter count to translate into a small download, the full 35B still has to be stored on disk
  • Anyone who needs benchmark claims independently verified before adopting a model, this released within days of when this page was written

So should you run it?

The architecture is byte-identical to Qwen 3.5 35B-A3B, the checkpoint Ornith AI continues-pretrains on top of, so the VRAM math carries over exactly. Q4_K_M puts the 256-expert weights at 21.3 GB; at an 8k-token working context that's a 24.1 GB total, already past a 24 GB card's real ~22.8 GB ceiling, so a 32 GB card (RTX 5090, RTX 5000 Ada) or a 32 GB+ unified-memory Mac is the comfortable target. The hybrid attention stack keeps a full 262,144-token session cheap once that weights floor is cleared: KV cache grows from 5.4 GB at native context to 29.9 GB in total, versus 21.5 GB of KV cache and 47.9 GB in total for a hypothetical all-attention version of the same model. Vision input needs the separate mmproj file, same packaging detail as every other model in this hybrid-stack family. Treat Ornith AI's own benchmark comparisons, including the Muse Glimmer 30B chart above, as vendor-reported until an independent group reproduces them; that caveat applies to any same-day model launch, not just this one.

Explore

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

Frequently asked questions

What are the VRAM requirements for Ornith 1.5 35B-A3B (MoE)?
Ornith 1.5 35B-A3B (MoE) requires approximately 24.1 GB of VRAM at Q4_K_M quantization, 41.9 GB at Q8, and 78.6 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 35B-A3B (MoE) have?
Ornith 1.5 35B-A3B (MoE) has 35 billion total parameters, but only 3 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
Can Ornith 1.5 35B-A3B (MoE) run on a 16 GB GPU?
No. At Q4_K_M, Ornith 1.5 35B-A3B (MoE) needs 24.1 GB of VRAM, more than 16 GB. You will need a 32 GB GPU like the RTX 5090.
Can Ornith 1.5 35B-A3B (MoE) run on a 24 GB GPU?
No. Even at Q4_K_M, Ornith 1.5 35B-A3B (MoE) needs 24.1 GB. Consider a 32 GB card like the RTX 5090.
What is the smallest quantization for Ornith 1.5 35B-A3B (MoE) that fits in 24 GB of VRAM?
At NVFP4, Ornith 1.5 35B-A3B (MoE) needs 19.8 GB, the highest-quality quantization that fits in 24 GB of VRAM.
What GPU do I need to run Ornith 1.5 35B-A3B (MoE) locally?
A 32 GB GPU is the minimum. At Q4_K_M, Ornith 1.5 35B-A3B (MoE) needs 24.1 GB VRAM. Good option: RTX 5090 (32 GB).
Is Ornith 1.5 35B-A3B the same model as Qwen 3.5 35B-A3B?
Architecturally, yes: config.json is byte-identical, 40 layers, 256 experts, the same hybrid Gated DeltaNet / Gated Attention stack. Ornith AI continues-pretrains on top of Alibaba's Qwen3.5 checkpoints (and, at other sizes, Google's Gemma 4) rather than training a new architecture, then applies its own self-improvement reinforcement-learning loop on top. The weights and the resulting benchmark scores are different releases, not the same download under two names.
How much better is this than the original Ornith 1.0 35B-A3B?
On the metric Ornith AI's own model card reports for both, Terminal-Bench 2.1 (Terminus-2 harness), Ornith 1.5 scores 67.8 against Ornith 1.0's 64.2, a modest but real gain from extending the self-improvement loop to also generate its own training tasks and scaffolds rather than only optimizing rollouts against a fixed task set.
Does Ornith 1.5 35B-A3B support image input?
The checkpoint carries Qwen3.5's vision tower (config.json's vision_config, an image and video token ID, and a matching mmproj GGUF file in the release), so the capability is present in the weights. Ornith AI's own documentation, however, describes this release purely as a text and coding-agent model and never demonstrates or claims multimodal use, so treat vision input as unverified rather than a documented feature.