Ornith 1.5 122B (MoE)
Ornith 1.5 122B (MoE) needs roughly 85.2 GB VRAM at Q4_K_M quantization (279.3 GB at FP16). 33 GPUs we track can run it fully in VRAM at 8k context.
33 GPUs run this natively · 11 with CPU offload
- Q4_K_M total
- 85.2 GB
- at 8k context
- Smallest GPU
- 96 GB
- NVIDIA RTX Pro 6000, at Q4_K_M
- KV cache, full context
- 6.4 GB
- 12 of 48 layers cache
- Inputs
- Text · Image · Video
- MIT, released October 2026
Ornith 1.5 122B (MoE) is a Mixture of Experts (MoE) model with 124.6B total parameters but only 10B active per token developed by Ornith AI. Added to the Hugging Face hub on 16 October 2026 under the MIT license, the mid-size Ornith-1.5 release after the 9B, 35B-A3B and 397B sizes of 19 August. A 256-expert mixture of experts (8 routed per token) with the same hybrid attention stack as Alibaba's Qwen 3.5 122B-A10B, whose config.json it matches field for field: 48 layers built from twelve repeats of three Gated DeltaNet layers followed by one Gated Attention layer, so only 12 of 48 layers keep a growing KV cache. The GGUF metadata puts the text model at 124.6B parameters. Ornith AI does not publish an active-parameter count; the 10B this site uses is carried over from the Qwen model with the same config. Context is 262,144 tokens natively and about 1M with YaRN scaling, which the card warns can slightly hurt short inputs. It is a reasoning model that opens each reply with a think block.
To run Ornith 1.5 122B (MoE) locally: The Q4_K_M file in ornith-ai's GGUF repo is 75.8 GB, and this site's estimate for the weights alone is 75.9 GB. Add a short context and the activation overhead and the total is 85.2 GB, past a single 80GB card even though the model card says it serves on one: that works at a lower quant, such as unsloth's 59.8 GB UD-Q3_K_XL, where this site's estimate is 67.3 GB in total. Context is the cheap part. The full 262,144-token window adds only 6.4 GB of KV cache, because 36 of the 48 layers carry a constant-size state instead of a cache that grows. A 96GB card holds Q4_K_M at a short context but not the full window (92.2 GB in total), so a 128GB unified-memory machine is the comfortable target for Q4_K_M at long context. As a MoE model, inference speed depends on active parameters (10B) rather than total size.
Ornith AI's own table reports SWE-bench Verified 84.5, SWE-bench Pro 76.9 and Terminal-Bench 2.1 81.2. The two SWE-bench rows are run with an anti-reward-hacking filter that marks any hacked solution wrong, and the Terminal-Bench row uses the Claude Code harness averaged over five runs. Against Qwen3.8-Flash-Next (125B, also tracked on this site) it leads on SWE-bench Pro (76.9 against 43.9) and SWE-bench Verified (84.5 against 82.8) but trails on Terminal-Bench 2.1 (81.2 against 83.4). These are Ornith AI's own runs, and no independent write-up of this size had appeared when this page was written.
VRAM at each quantization
Numbers here are computed at 8k context. This model's hybrid attention stack means KV cache grows much more slowly than context length, unlike a conventional full-attention model.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 498.4 GB | 0.20 GB | 558.4 GB |
| BF16 | 249.2 GB | 0.20 GB | 279.3 GB |
| FP16 | 249.2 GB | 0.20 GB | 279.3 GB |
| Q8_0 | 132.4 GB | 0.20 GB | 148.6 GB |
| Q6_K | 102.3 GB | 0.20 GB | 114.8 GB |
| Q5_K_M | 88.7 GB | 0.20 GB | 99.6 GB |
| Q4_K_Mrec | 75.9 GB | 0.20 GB | 85.2 GB |
| Q3_K_M | 59.9 GB | 0.20 GB | 67.3 GB |
| Q2_K | 47.5 GB | 0.20 GB | 53.4 GB |
| NVFP4cuda | 77.6 GB | 0.20 GB | 87.1 GB |
KV cache figures assume 8k context at FP16. NVFP4 quantization requires a CUDA-capable GPU. Enable TurboQuant in the calculator to see reduced KV cache estimates.
Benchmarks
GPUs that run Ornith 1.5 122B (MoE) natively (33)
- NVIDIA B300 288GBNVFP4 · 136.6 t/s
- NVIDIA B200 180GBNVFP4 · 136.6 t/s
- NVIDIA H200 141GBQ6_K · 109.1 t/s
- NVIDIA H100 80GBQ3_K_M · 132.2 t/s
- NVIDIA A100 80GBQ3_K_M · 115.5 t/s
Show 28 more
- NVIDIA RTX Pro 6000NVFP4 · 124.8 t/s
- NVIDIA DGX Spark (128GB)NVFP4 · 32.6 t/s
- AMD Instinct MI300XQ8_0 · 66.1 t/s
- AMD Strix Halo (128GB)Q6_K · 21.7 t/s
- AMD Strix Halo (96GB)Q4_K_M · 27.3 t/s
- AMD Strix Halo (64GB)Q2_K · 37.8 t/s
- Apple M5 Ultra (512GB)BF16 · 31.6 t/s
- Apple M5 Ultra (256GB)Q8_0 · 45.8 t/s
- Apple M5 Ultra (96GB)Q4_K_M · 58.4 t/s
- Apple M5 Max (128GB)Q6_K · 36.1 t/s
- Apple M5 Max (64GB)Q2_K · 53.4 t/s
- Apple M5 Pro (64GB)Q2_K · 37.2 t/s
- Apple M4 Max (128GB)Q6_K · 33.5 t/s
- Apple M4 Max (64GB)Q2_K · 50.6 t/s
- Apple M3 Ultra (512GB)BF16 · 24.2 t/s
- Apple M3 Ultra (256GB)Q8_0 · 36.9 t/s
- Apple M3 Ultra (96GB)Q4_K_M · 49.6 t/s
- Apple M3 Max (128GB)Q6_K · 27.1 t/s
- Apple M3 Max (96GB)Q4_K_M · 27.2 t/s
- Apple M3 Max (64GB)Q2_K · 43.3 t/s
- Apple M2 Ultra (192GB)Q8_0 · 36.4 t/s
- Apple M2 Ultra (64GB)Q2_K · 59.4 t/s
- Apple M2 Max (96GB)Q4_K_M · 33.1 t/s
- Apple M2 Max (64GB)Q2_K · 43.3 t/s
- Apple M1 Ultra (128GB)Q6_K · 42.2 t/s
- Apple M1 Ultra (64GB)Q2_K · 59.4 t/s
- Apple M1 Max (64GB)Q2_K · 43.3 t/s
- Intel Data Center GPU Max 1550Q6_K · 67.6 t/s
Plus 11 GPUs that run it with CPU offload (slower)
- NVIDIA RTX 5090Q2_K · 24.8 t/s
- NVIDIA A100 40GBQ2_K · 32.5 t/s
- NVIDIA L40SQ3_K_M · 23 t/s
- NVIDIA RTX A6000Q3_K_M · 23.5 t/s
- NVIDIA RTX 5000 AdaQ2_K · 22.5 t/s
- NVIDIA RTX 6000 AdaQ3_K_M · 24.2 t/s
- AMD Radeon PRO W7800Q2_K · 20.3 t/s
- AMD Radeon PRO W7900Q3_K_M · 21.4 t/s
- AMD Radeon AI PRO R9700 32GBQ2_K · 20.6 t/s
- Intel Arc Pro B70 32GBQ2_K · 20.5 t/s
- Intel Data Center GPU Max 1100Q3_K_M · 21.5 t/s
Cloud GPU Rental
Cheapest way to run Ornith 1.5 122B (MoE): 1× RTX Pro 6000 at NVFP4, ~125 tok/s, from $1.69/hr.
Pay by the hour · no contract · pods start in about a minute.
Rent this setup on RunPod ↗ (+$5 signup credit)Prices as of Oct 17, 2026. Rates change; check RunPod for current pricing.
Affiliate link: CanItRun may earn a commission. Doesn't affect the fit calculation above.
Notes
Added to the Hugging Face hub on 16 October 2026, after the 9B, 35B-A3B and 397B sizes of 19 August. A 256-expert MoE with the same hybrid attention layout as Alibaba's Qwen 3.5 122B-A10B; Ornith AI publishes no active-parameter count, so the 10B here is carried over from that model. Ornith AI's own table leads Qwen3.8-Flash-Next (125B) on SWE-bench Pro (76.9 against 43.9), SWE-bench Verified (84.5 against 82.8) and NL2Repo (51.2 against 45.2) and trails it on Terminal-Bench 2.1 (81.2 against 83.4, Claude Code harness). These are Ornith AI's own runs, and no independent write-up of this size had appeared when this page was written. Builds from ornith-ai: GGUF (Q4_K_M 75.8 GB, Q5_K_M 88.7 GB, Q6_K 102.4 GB, Q8_0 132.6 GB), FP8 (132.9 GB), NVFP4 (77.6 GB) and MLX 4-bit (68.7 GB) and 8-bit (129.7 GB); unsloth adds dynamic GGUFs from 40.4 GB. MIT licensed.
Compare Ornith 1.5 122B (MoE) with other models
How to run Ornith 1.5 122B (MoE) locally
Q4_K_M needs 85.2 GB: needs multiple datacenter-class GPUs (80 GB+ each).
llama.cpp (Ornith AI's recommended sampling), or vLLM/SGLang on an 80GB-class GPU
./llama-server \
-hf ornith-ai/Ornith-1.5-122B-GGUF:Q4_K_M \
-c 262144 -ngl 99 \
--temp 0.6 --top-p 0.95 --top-k 20Ollama
ollama run ornith-1.5:122bllama.cpp
# General tasks (Ornith AI's other profile: temp 1.0, presence penalty 1.5)
./llama-server \
-hf ornith-ai/Ornith-1.5-122B-GGUF:Q4_K_M \
-c 262144 -ngl 99 \
--temp 1.0 --top-p 0.95 --top-k 20 --presence-penalty 1.5
# Vision needs the projector file alongside the weights
./llama-server -hf ornith-ai/Ornith-1.5-122B-GGUF:Q4_K_M \
--mmproj mmproj-Ornith-1.5-122B-BF16.ggufLM Studio: Search for 'Ornith 1.5 122B' and take the Q4_K_M build from ornith-ai, or one of unsloth's smaller dynamic quants if you are short on memory. The model is a reasoning model, so expect a think block before every answer. Most of the memory goes to the weights rather than the context, so budget for the 75.8 GB file first and treat the context slider as the smaller cost.
Why this quantization? Q4_K_M puts the 256-expert weights at roughly 75.9 GB by this site's estimate (75.8 GB as published), which is more than a single 80GB card can hold once activation overhead and a working context are added. A 96GB card runs it at a short context (85.2 GB in total at 8k) but not at the full window (92.2 GB), and a 128GB unified-memory machine runs it at both. The hybrid attention stack keeps context cheap once that floor is cleared: the full 262,144-token window adds only about 6.4 GB of KV cache, since just 12 of 48 layers grow their cache with context. On an 80GB card, drop to a 3-bit dynamic quant such as unsloth's UD-Q3_K_XL (59.8 GB) rather than squeezing Q4_K_M.
Who is Ornith 1.5 122B (MoE) for?
Anyone with a 96GB workstation card (short context) or a 128GB unified-memory machine (full context) who wants an open-weight coding-agent model in the 120B class. Ornith AI's own table has it ahead of Qwen3.8-Flash-Next on SWE-bench Pro and Verified, with about 10B parameters active per token keeping decode fast.
Best for
- Local coding agents that need more headroom than the 35B-A3B and can spare roughly 80 to 100GB of memory
- Long-document and multi-file analysis, where the hybrid attention stack keeps KV cache growth flat
- Terminal and tool-use agents through an OpenAI-compatible server with the qwen3_xml tool parser
- Image understanding via the vision projector
Not ideal for
- Any consumer GPU of 32GB or less, which would have to hold most of the 75.8 GB of Q4_K_M weights in system RAM
- Anyone who needs the roughly 10B active-parameter count to translate into a small download, all 124.6B parameters still have to be stored
- Anyone who needs benchmark claims independently verified before adopting a model, this was added to the Hub within days of when this page was written and every comparison on the card is Ornith AI's own
Does it fit on one 80GB card?
Not at Q4_K_M. The weights alone are 75.9 GB, and at an 8k-token working context this site's total is 85.2 GB, past an 80 GB card; the full 262,144-token window brings it to 92.2 GB, because the KV cache adds only 6.4 GB. A 96 GB workstation card holds Q4_K_M at a short context but runs out of usable memory just before the full window, while a 128 GB unified-memory machine holds both. Ornith AI's model card says the model serves on a single 80 GB GPU without naming a quantization, and since the BF16 weights alone are about 249 GB it can only mean a quantized build. A 3-bit build is the realistic fit, at 59.9 GB of weights and 67.3 GB in total at 8k context. Ornith AI's own NVFP4 checkpoint is 77.6 GB, or 87.1 GB in total at 8k context, so it needs the same 96 GB class of card as Q4_K_M and has the same limit at long context. Every benchmark figure on this page is Ornith AI's own, so treat the coding numbers as a claim to test on your own repository.
Frequently asked questions
- What are the VRAM requirements for Ornith 1.5 122B (MoE)?
- Ornith 1.5 122B (MoE) requires approximately 85.2 GB of VRAM at Q4_K_M quantization, 148.6 GB at Q8, and 279.3 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 122B (MoE) have?
- Ornith 1.5 122B (MoE) has 124.6 billion total parameters, but only 10 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 122B (MoE) run on a 16 GB GPU?
- No. At Q4_K_M, Ornith 1.5 122B (MoE) needs 85.2 GB of VRAM, more than 16 GB. You will need a multi-GPU server.
- Can Ornith 1.5 122B (MoE) run on a 24 GB GPU?
- No. Even at Q4_K_M, Ornith 1.5 122B (MoE) needs 85.2 GB. Consider a multi-GPU server with 86 GB+ of combined VRAM.
- What is the smallest quantization for Ornith 1.5 122B (MoE) that fits in 24 GB of VRAM?
- Ornith 1.5 122B (MoE) cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 53.4 GB at Q2_K.
- What GPU do I need to run Ornith 1.5 122B (MoE) locally?
- You need a multi-GPU server. At Q4_K_M, Ornith 1.5 122B (MoE) needs 85.2 GB VRAM, more than any single consumer GPU. That's roughly 2x 80 GB datacenter GPUs (H100, A100, or similar) pooled together.
- Can I run Ornith 1.5 122B (MoE) in the cloud?
- Yes. The cheapest RunPod setup we found that holds Ornith 1.5 122B (MoE) entirely in GPU memory at 8k context is 1× RTX Pro 6000 at NVFP4, at roughly 124.8 tokens per second and from $1.69 per hour for the whole pod. Prices as of Oct 17, 2026; RunPod rates change, so check their pricing page for the current figure.
- Can Ornith 1.5 122B run on a single 80GB GPU?
- Only at a lower quant. At Q4_K_M this site's total is 85.2 GB at an 8k-token context, more than an H100 or A100 80GB holds, while a 3-bit build comes to 67.3 GB and fits with room for context. Ornith AI's model card says the model serves on a single 80 GB GPU but does not say which quantization; the BF16 weights alone are about 249 GB, so it has to be a quantized build. A 96 GB card such as the RTX Pro 6000 holds Q4_K_M at 8k context but not the full 262,144-token window (92.2 GB in total); a 128 GB unified-memory machine holds both.
- Is Ornith 1.5 122B the same model as Qwen 3.5 122B-A10B?
- Architecturally, yes: config.json matches field for field, 48 layers, 256 experts with 8 routed per token, 3,072 hidden size, and 12 full-attention layers that keep a KV cache, so the memory profile is the same. The other three Ornith-1.5 sizes are built on Qwen3.5 checkpoints; the 122B card does not name its base model. Any difference in quality comes from the training, and the benchmark figures in the comparison are Ornith AI's own.
- How much memory does the long context cost on Ornith 1.5 122B?
- Very little next to the weights: the full 262,144-token window adds 6.4 GB of KV cache, because only 12 of the 48 layers keep one and the other 36 carry a constant-size state. If all 48 layers cached, the same window would need 25.8 GB. The extended 1M-token setting uses YaRN scaling at factor 4.0, which the model card warns can slightly hurt quality on short inputs.