What it is
Ornith-1.5-122B is the 122-billion-parameter member of Ornith's 1.5 family. Its weights appeared on Hugging Face on 16 October 2026, with FP8, NVFP4, GGUF and MLX builds following on 20 October, two months after the 9B, 35B and 397B sizes. Its configuration uses Qwen3.5's mixture-of-experts architecture (256 experts, 8 routed per token, 48 layers mixing linear and full attention); Ornith does not publish the active parameter count. Training extends Ornith-1.0, itself built on Qwen3.5 and Gemma 4: the model proposes new training tasks, builds the agent setup to attempt each one, and learns from the results through reinforcement learning.
Ornith reports 84.5% on SWE-bench Verified, 76.9% on SWE-bench Pro and 81.2% on Terminal-Bench 2.1, scored with what it calls anti-hacking safeguards, and claims leads over Qwen3.8-Flash-Next and Nemotron 3.5 Super VL on most coding and agentic tests. It reasons before answering, takes text and images, supports tool calling and a 262,144-token context, and runs on vLLM and SGLang. The card says it serves on a single 80 GB GPU, but only the 4-bit builds (about 76 to 78 GB) come close; FP8 is 133 GB and BF16 250 GB.
The weights are under the MIT licence.
Our review
Ornith-1.5-122B posts the most striking numbers in its weight class, and the claim that matters most to buyers is weaker than it looks. The model card says it serves on a single 80 GB GPU, yet the BF16 checkpoint is 250 GB and the FP8 build 133 GB. Only the 4-bit releases, NVFP4 at 77.7 GB and the Q4_K_M GGUF at 75.8 GB, come near one 80 GB card, and they leave little room for the cache a 262,144-token context needs.
On Ornith's own table it scores 76.9% on SWE-bench Pro against 43.9% for Qwen3.8-Flash-Next, a 33-point gap that calls for independent confirmation. Elsewhere it is closer: Qwen3.8-Flash-Next leads Terminal-Bench 2.1 at 83.4% to 81.2%, and on SWE-bench Verified the margin is under two points. Its "no-hack" setting means models were warned about reward hacking and any solution that still did it was marked wrong; that is more care than most labs show, but Ornith wrote the rules, applied them, and used Claude Opus as the judge on two benchmarks.
The training method is the reason to watch the lab. Each cycle has the model propose harder tasks, scored for validity, novelty and difficulty around a 20% target success rate, then build the scaffold to attempt them and learn from the attempts with GRPO, a reinforcement-learning method. The blog names a hack-resistance term in the reward but never says how it is computed, which matters for a loop that writes its own exams.
Adoption is easy for anyone already on Qwen. The architecture is Qwen3.5's, the serve commands use Qwen's tool and reasoning parsers, and vLLM 0.19.1+, SGLang 0.5.9+ and Ollama all work. Ornith publishes no people, location or funding, yet its 35B GGUF drew about 3.9 million Hugging Face downloads in the past month, so the family has real users even if the lab stays anonymous. The August blog post that explains the method predates the 122B and does not cover it.
Where it earns its place
An ML platform engineer with one 80 GB GPU
Replace a Qwen3.5-based coding model without rewriting the serving stack
Swap the model path in vLLM or SGLang and keep the qwen3 reasoning and tool parsers. In practice that means the 4-bit build and a context well below 262K to leave memory for cache; run your own regression set first, since every score is Ornith's.
An infrastructure team running a two-GPU 80 GB node
Serve an internal coding agent at FP8 rather than 4-bit quality
The 133 GB FP8 checkpoint needs both cards with tensor parallelism set to 2, which doubles the hardware the card's headline implies. Prefix caching in the documented vLLM command helps with repeated repository context.
A head of engineering at a regulated fintech
Run an agentic coding model on premises under a permissive licence
MIT weights allow commercial use and fine-tuning without negotiation, and code never leaves the network. The trade-off is supplier risk: no named team, support contract or first-party API to fall back on.
Before you commit
- Setup
- `ollama run ornith-1.5:122b` pulls 73 to 77 GB. For servers, the card's vLLM or SGLang commands need Transformers 5.8.1+, vLLM 0.19.1+ or SGLang 0.5.9+.
- Integrations
- OpenAI-compatible chat completions with tool calls and a separate reasoning_content field; the card shows Ollama, llama.cpp, Atomic.chat and Hermes Agent. Image input uses the mmproj file in the GGUF repository.
- Data and privacy
- Self-hosted only, so data stays wherever you run it. There is no first-party API, and so no retention or training terms to review.
- Cost at scale
- Weights free. BF16 is 250 GB (a multi-GPU node), FP8 133 GB (two 80 GB cards), and NVFP4 (77.7 GB) or Q4_K_M (75.8 GB) the only builds near one 80 GB card; Macs get a 68.7 GB MLX 4-bit build. No first-party hosted price published; NanoGPT lists $0.20 input and $0.80 output per million tokens but marks the model unavailable.
- Limits
- Reasoning is on by default, so answers follow a think block. Context is 262,144 tokens, extendable to about 1M with YaRN at some cost to quality on normal-length inputs.
What's good
- Reported 84.5% SWE-bench Verified, 76.9% Pro
- MIT licence
- Drop-in for Qwen3.5 vLLM or SGLang stacks
- FP8, NVFP4, GGUF and MLX builds from launch
What's not
- "Single 80 GB GPU" only at 4-bit
- All scores self-reported, under rules Ornith set
- No named team, funding or first-party API
- Trails Qwen3.8-Flash-Next on Terminal-Bench 2.1
Best for
Teams serving open-weight coding agents on a single large GPU who can run their own evaluations.
Alternatives
Ornith-1.5-122B: common questions
- Is Ornith-1.5-122B worth it?
- Striking coding scores from a lab with no public names, and a one-GPU pitch that holds only at 4-bit. We rate it 3.6 out of 5.
- How much does Ornith-1.5-122B cost?
- Free open weights under MIT. You pay only for your own hardware or hosting.
- What are the downsides of Ornith-1.5-122B?
- "Single 80 GB GPU" only at 4-bit; all scores self-reported, under rules Ornith set; no named team, funding or first-party API; trails Qwen3.8-Flash-Next on Terminal-Bench 2.1.
- Who is Ornith-1.5-122B best for?
- Teams serving open-weight coding agents on a single large GPU who can run their own evaluations.
- How much does Ornith-1.5-122B cost for a team or in production?
- Weights free. BF16 is 250 GB (a multi-GPU node), FP8 133 GB (two 80 GB cards), and NVFP4 (77.7 GB) or Q4_K_M (75.8 GB) the only builds near one 80 GB card; Macs get a 68.7 GB MLX 4-bit build. No first-party hosted price published; NanoGPT lists $0.20 input and $0.80 output per million tokens but marks the model unavailable.
- Is Ornith-1.5-122B safe for company data?
- Self-hosted only, so data stays wherever you run it. There is no first-party API, and so no retention or training terms to review.
- What are the best Ornith-1.5-122B alternatives?
- The closest alternatives we have reviewed are Qwen3.8-Flash-Next, Nemotron 3.5 Super VL and DeepSeek-V4.1-Flash.
Answers checked Oct 23, 2026.