Ornith Drops Ornith-1.5 122B Open Coding Model Claiming 84.5 on SWE-bench
A 122B mixture-of-experts reasoning model with MIT weights, a 262K context, and 84.5% on SWE-bench Verified, served from a single 80GB GPU.
- Ornith released Ornith-1.5-122B, a 122B MoE reasoning model under MIT license.
- Fits on a single 80GB GPU; GGUF, FP8, NVFP4, MLX, and Mobile variants shipped day one.
- Reports 84.5% SWE-bench Verified, 76.9% SWE-bench Pro no-hack, 81.2 Terminal-Bench 2.1.
- Native 262K context, extendable to ~1M tokens via YaRN rope scaling.
- Trained via self-improvement loop: model generates tasks, scaffolds, and RL rollouts itself.
- Hosted pricing around $0.26 in / $2.08 out per million tokens on third-party providers.
Ornith has released Ornith-1.5-122B, an open-weight mixture-of-experts model aimed at coding agents, terminal work, and multi-step tool use. The group reports 84.5 on SWE-bench Verified, 76.9 on SWE-bench Pro under its “no-hack” setup, and 81.2 on Terminal-Bench 2.1. It also claims a quantized build can run on one 80GB NVIDIA GPU.
If independent evaluations reproduce those results, the release would give teams a locally deployable alternative for coding workflows that currently depend on hosted models. The single-GPU claim requires qualification, however: the smallest listed GGUF file occupies 75.8GB before runtime overhead and context memory are included.
The GGUF repository has recorded more than 850,000 downloads. Hugging Face counts file requests, including repeated and automated downloads, so that figure does not represent distinct users or production deployments.
The release at a glance
| Item | Reported detail |
|---|---|
| Architecture | 122B-parameter mixture of experts, 48 layers |
| Expert routing | 256 routed experts, top-8 routing, plus one shared expert |
| Attention | 3:1 mix of linear GDN blocks and full attention |
| Native context | 262,144 tokens |
| Extended context | Approximately 1 million tokens with YaRN scaling |
| Other components | One multi-token-prediction layer and a vision tower |
| License | MIT, according to the repository |
| Minimum runtimes | Transformers 5.8.1, vLLM 0.19.1, or SGLang 0.5.9 |
A mixture-of-experts model routes each token through a subset of its experts, reducing computation relative to a dense model of the same total size. All expert weights still need storage, and the release does not state the number of active parameters per token. That missing figure makes throughput difficult to estimate from the 122B total alone.
The native 262K context is distinct from the roughly 1M-token YaRN mode. YaRN extends positional scaling beyond the trained context window, which can increase reach while changing quality and memory requirements. Long-context accuracy needs separate testing at the lengths an application will use.
Weights for servers, desktops, and Macs
The repository declares an MIT license and provides several weight formats for different inference stacks:
- BF16: approximately 249GB, requiring multiple GPUs or substantial unified memory.
- FP8 and NVFP4: lower-precision variants for supported accelerators.
- GGUF: llama.cpp-compatible quantizations ranging from Q4_K_M at 75.8GB to Q8_0 at 133GB.
- MLX: 4-bit and 8-bit builds for Apple Silicon.
- Mobile: a separately quantized Ornith-1.5-122B-Mobile variant aimed at constrained devices.
Teams deriving the model or redistributing it should verify the exact upstream checkpoints, license notices, and attribution requirements. The release description refers to additional training applied to Qwen3.5 and Gemma4, while describing the 122B architecture as Qwen3.5-style.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
