Ornith-1.5-122B Ships MIT Weights With 10B Active Parameters and an 84.5 SWE-bench Verified Score, but It Fits One 80GB GPU Only at 4-Bit and Its Own Table Shows Rivals Ahead on 6 of 14 Tests

Ornith released Ornith-1.5-122B on 20 October 2026, a mixture-of-experts coding and agent model with open weights on Hugging Face and a same-day Ollama tag. The model card headlines an 84.5 on SWE-bench Verified and says the model “serves on a single 80GB GPU.”

The launch spread fast on X, with community quants, free trials on two hosted services and day-zero SGLang support all landing within hours. We went through the Ornith-1.5-122B open weights repo file by file: the benchmark table, the config, every official build size and the license link.

The short version: the coding numbers are strong for a 10B-active model, but the “one 80GB GPU” line only holds at 4-bit, and Ornith’s own table shows a rival ahead on 6 of its 14 rows.

What Ornith actually shipped

Ornith-1.5-122B joins the Ornith-1.5 family that launched in August with 9B, 35B and 397B models. According to the Ornith-1.5 blog post, the family extends Ornith-1.0, which was built on top of Qwen3.5 and Gemma 4 with extra pretraining and post-training. The 122B card repeats that lineage.

The config file lists a Qwen3.5 MoE architecture with 48 layers, 256 experts and 8 experts routed per token, plus a shared expert. When a user asked about active parameters, Ornith replied on X that “It has 10B active parameters.”

The weights total 125.09 billion parameters in BF16, not 122 billion. The index file shows why: alongside the language model it ships a vision encoder and a multi-token prediction layer, which is also why Ollama lists text and image input.

The benchmark table, row by row

Ornith compares the 122B model with Qwen3.8-Flash-Next (125B), Qwen3.8 (27B dense), Nemotron 3.5 Super VL (120B), DeepSeek-V4-Flash-0731 (284B) and Reflection’s Beam (501B). All scores come from Ornith’s own runs.

Ornith leads on the three SWE-bench variants, CyberGym, HLE without tools, SciCode and ClawEval. It trails on TerminalBench, NL2Repo, HLE with tools, WideSearch, BrowseComp and Toolathlon. Several rival cells are blank, so the SWE-bench wins are against fewer competitors than the table width suggests.

The SWE-bench Pro gap stands out. Ornith scores 76.9 while Qwen3.8-Flash-Next, the closest model by size, gets 43.9 in the same table. Ornith says it applies anti-hacking rules and marks any solution that still shows reward-hacking behaviour as incorrect, so how each model was scored under that rule matters. No independent run has been published yet.

Two more details for careful readers. HLE was graded with Claude 4.6 Opus as the judge, and MCP-Atlas with Claude 4.8 Opus. And the chart image in the repo shows DeepSeek and Beam at 35 and 36 on HLE, while the text table lists them at 50.8 and 51.1 with tools and gives no score without tools, so the chart and table do not fully match. Ornith’s launch post on X gives a third HLE figure, 0.446, which matches neither the 48.0 with tools nor the 37.68 without tools on the card.

The DeepSeek comparison is also dated. The table uses the July V4-Flash build, not the DeepSeek V4.1 Flash 1017 release that shipped three days earlier.

Does it really fit on one 80GB GPU?

This is where the card is most generous. It says the model “serves on a single 80GB GPU,” then gives a vLLM command pointing at the full BF16 repo. That repo is 250.2 GB. The model card also says it “runs on a single DGX,” and the Ollama page says “a single NVIDIA DGX card,” which most posters on X read as NVIDIA’s DGX Spark desktop with 128 GB of unified memory.

We added up every official build from the Hugging Face file listings:

So on a single 80GB card such as an H100, only the 4-bit builds fit, and they leave a few gigabytes for the KV cache. That is far short of the 262K context the serve commands set. The NVFP4 build also needs a Blackwell-class GPU to run natively.

Community testers fill in the rest. One user reported the Q4_K_M GGUF on a DGX Spark at 28.2 tokens a second for a single user, with about 80 GiB of memory in use. Unsloth co-founder Daniel Han posted a 2-bit dynamic quant at 44.5 GB. These are individual reports, not Ornith claims.

The card also mentions an “Ornith-1.5-122B-Mobile” quantized variant for phones. That repo was not public on Hugging Face when we checked on 21 October.

Price and access

There is no paid Ornith API listed for this model, and it was not on OpenRouter’s model list when we checked. The weights are free to download, so the real cost is your hardware or GPU rental.

For a hosted trial, Ornith said the model is free on Command Code “for a limited time,” and Tommy of Nous Research, in a post Ornith reposted, said it is free in Nous Portal for 48 hours. Ornith also announced SGLang support. Serving needs vLLM 0.19.1, SGLang 0.5.9 or Transformers 5.8.1 or newer, per the card.

Confirmed vs not yet shown

On the license: the card metadata says MIT and links to a LICENSE file, but that link returned “Entry not found” when we checked, and the repo file list has no LICENSE file. If you plan commercial use, keep a copy of the card as it reads today and watch for the file to appear.

What it means for Indian developers

A 10B-active model with MIT weights is a practical option for teams that cannot send code to a US API, for example under client contracts or India’s Digital Personal Data Protection Act. You can run it inside your own cloud account or office.

The catch is memory. Plan for a 128 GB class machine, such as a DGX Spark, a high-memory Mac for the 68.7 GB MLX 4-bit build, or two 80GB cloud GPUs if you want long context at higher precision. A single rented 80GB GPU works for short prompts at 4-bit, not for repo-scale agent sessions.

If your workload is mostly SWE-bench style bug fixing, Ornith’s own numbers favour it. If it leans on terminal work or web research, the same table points to Qwen3.8-Flash-Next or DeepSeek. Test on your own repos before switching, as we suggested with Reflection’s Beam open weights and Meta’s Muse Spark Open.

FAQ

What is Ornith-1.5-122B?

It is an open-weight mixture-of-experts reasoning model from Ornith, released on 20 October 2026. It has about 125B total parameters including vision and MTP weights, 10B active per token, and a 262K-token native context.

Can Ornith-1.5-122B run on a single GPU?

Only at 4-bit on an 80GB card. The NVFP4 build is 77.7 GB and the Q4_K_M GGUF is 75.8 GB, which leaves little room for context. A 128 GB DGX Spark can hold the Q6_K and Q5_K_M builds.

How does it compare with Qwen3.8-Flash-Next and DeepSeek?

In Ornith’s own table it leads on SWE-bench Verified, Pro and Multilingual, but trails Qwen3.8-Flash-Next on TerminalBench and Toolathlon and DeepSeek-V4-Flash on NL2Repo, WideSearch and BrowseComp.

How much does Ornith-1.5-122B cost?

The weights are free to download. No paid API price has been published; Command Code and Nous Portal offered free limited-time access at launch.

Is it really MIT licensed?

The Hugging Face card metadata says MIT, but the LICENSE file it links to was missing from the repo on 21 October 2026. Check the repo again before relying on it for commercial work.