Ornith-1.5-397B
MoE workstation~403B total params, MoE - Ornith AI’s flagship self-improving agentic-coding model, released 2026-08-18. Built on continued pretraining of Qwen 3.5 / Gemma 4, then a self-improvement loop where the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for RL (all three stages optimized jointly with GRPO). 262K context, MIT-licensed on HuggingFace at ornith-ai/Ornith-1.5-397B (FP8, GGUF, and NVFP4 quantizations).
- Coding (vendor self-reported, 5-run average): Terminal-Bench 2.1 86.1, SWE-bench Verified 86.0, SWE-bench Pro 65.1, DeepSWE 56.0, NL2Repo 59.5 - on par with Claude Opus 4.8 and ahead of GLM-5.2 and DeepSeek-V4-Flash-0731.
- Reasoning: HLE 44.6 (no tools) / 56.1 (with tools), GPQA-Diamond 92.8.
- Agentic: MCP-Atlas 80.0, Toolathlon-Verified 71.2, ClawEval 81.4.
Server-class only. ~403B MoE needs serious multi-GPU hardware (or a quantized GGUF on a workstation-class machine); it is not a single-RTX-5090 or DGX Spark model. Treat vendor benchmarks as claims until independently replicated on Artificial Analysis.
- 403.0B
- 262k
- mit
- 🇺🇸 USA
- Aug 2026
Benchmark scores
Vendor-reported - from the developer's own model card / tech report
Vendor-reported - from the developer's own model card / tech report
Ran this model on your own hardware? Join free and add your measured tok/s to the community numbers.
Related models
Save your hardware and every model page answers the real question: will it run on your machine, and how fast?
Join free - save your rig →Run it locally
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
The reference hardware
22 reference configs, drawn in-house. Scroll for more.
Can you run it? - reference rigs
| Rig | Q4_K_M | Q5_K_M | Q6_K | Q8_0 |
|---|---|---|---|---|
| NVIDIA Jetson Orin NX 16GB | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| Jetson AGX Orin 64GB | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| Single GTX 1080 Ti (11GB) | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| Single RTX 4090 (24GB) | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| Single RTX 5090 (32GB) | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| RTX PRO 6000 Blackwell (96GB) | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| 2x RTX 3090 (48GB) | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| 2x RTX 5090 (64GB) | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| 4x RTX 4090 (96GB) | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| 4x RTX 5090 (128GB) | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| 8x RTX 3090 rack (192GB) | offload | offload | no -> cloud | no -> cloud |
| AMD Instinct MI300X (192GB) | offload | offload | no -> cloud | no -> cloud |
| MacBook Pro M5 Max 128GB | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| Ryzen AI Max+ 395 128GB | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| Mac Studio M4 Ultra 192GB | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| DGX Spark 128GB unified | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| 4x H100 80GB (320GB) | fast 29.2t/s | fast 25.1t/s | offload | offload |
| NVIDIA DGX Station 748GB | ok 17.5t/s | ok 15.0t/s | ok 13.0t/s | ok 10.1t/s |
| Mac Studio M4 Ultra 512GB | slow 2.6t/s | slow 2.2t/s | slow 1.9t/s | slow 1.5t/s |
| Dual EPYC 9004 + 768GB DDR5-4800 | slow 1.0t/s | slow 0.9t/s | slow 0.8t/s | slow 0.6t/s |
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 | slow 0.5t/s | slow 0.4t/s | slow 0.3t/s | slow 0.3t/s |
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 | slow 0.3t/s | slow 0.3t/s | slow 0.3t/s | slow 0.2t/s |
Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
Download options
Or run it in the cloud
No per-token API provider pricing tracked for Ornith-1.5-397B yet. For flagship list prices, see the calculator.
Inference cost over time
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.