How does Ornith-1.5-122B score against Qwen 3.8 and DeepSeek V4 Flash?

How does Ornith-1.5-122B score against comparable open models?

On its model card, Ornith-1.5-122B leads or ties most coding and agentic benchmarks against Qwen3.8-Flash-Next (125B), Qwen3.8 27B, and Nemotron 3.5 Super VL (120B). Its biggest result is SWE-bench Pro (no-hack), where it scores 76.9 against 65.5 for the next-best listed model. Against the larger DeepSeek-V4-Flash-0731 (284B) and Beam (501B), it trades wins. It trails DeepSeek on several search and reasoning tests. All numbers are self-reported.

What is Ornith-1.5-122B?

Ornith-1.5-122B is the largest-capability member of the Ornith-1.5 family. The card describes it as a 122B mixture-of-experts model designed for single-GPU deployment, and says it runs on a single DGX. A quantized Ornith-1.5-122B-Mobile variant targets edge devices.

The team says 1.5 extends Ornith-1.0 by widening a self-improvement loop. Earlier work optimized scaffolds and rollouts. This version jointly optimizes task generation, scaffold construction, and solution rollouts. The pitch is "stronger, faster, leaner": better scores, quicker task completion, and shorter outputs. The card gives no table for speed or output length, so only the benchmark claims can be checked against the published numbers.

A reviewer who tested the model locally also reports that it reads images, reasons before answering, and handles a context window of just over 260,000 tokens.

How does it score on coding benchmarks?

Coding is where the card makes its strongest case.

Ornith tops CyberGym and all three SWE-bench variants where rivals have scores. It loses Terminal-Bench 2.1 to both Qwen3.8-Flash-Next and DeepSeek-V4-Flash, and NL2Repo to DeepSeek-V4-Flash.

The SWE-bench Pro row stands out. Ornith is 33 points above Qwen3.8-Flash-Next, a model of nearly the same size. The 27B dense Qwen3.8 scores 65.5, yet the larger Qwen3.8-Flash-Next scores 43.9. That inversion suggests the setup, scaffold, or settings differ across entries. Treat it as a reason to rerun the test, not a settled ranking.

How does it score on reasoning and agentic tests?

The reasoning picture is more mixed.

Among the same-size models, Ornith leads every reasoning row. With tools on HLE, both DeepSeek-V4-Flash and Beam finish ahead of it by a few points.

Agentic results look like this:

The ClawEval row is cut off in the source I had, so only Ornith and Qwen3.8-Flash-Next appear. The card also lists further evaluations beyond this point that I could not read.

On agentic tasks the margins are small. Ornith ties DeepSeek-V4-Flash on MCP-Atlas at 74.6, edges it on Toolathlon-Verified by 0.5, and loses to it on WideSearch and BrowseComp. Qwen3.8-Flash-Next wins Toolathlon-Verified by 0.6. Differences under a point or two on agent benchmarks are well within normal run-to-run noise.

Does Ornith really match models two to four times its size?

Partly. DeepSeek-V4-Flash-0731 has 284B parameters and Beam has 501B. Ornith beats them on some coding rows (CyberGym, and SWE-bench Pro against Beam) and ties DeepSeek on MCP-Atlas, which is an agentic row rather than a coding one. It sits a few points behind on search-heavy and tool-assisted reasoning tests. Many cells for those larger models are empty, so a full head-to-head is not possible from this table.

Parameter count is also a loose comparison for mixture-of-experts models. Total parameters say little about active compute per token, and the card does not list an active-parameter figure.

What caveats should you keep in mind?

• Self-reported scores. The comparison table sits on the vendor's own card. Rival numbers may come from their publishers, from the Ornith team's reruns, or both. The card does not say.
• Undefined labels. "No-hack" and "non-hack" appear on the SWE-bench rows without explanation in the part of the card I could read. They presumably mean the runs excluded known exploits, but verify before relying on that reading.
• Missing cells. Several rivals have dashes, especially on SWE-bench Verified and Multilingual, so "leads the chart" can mean "leads the models that were scored."
• Scaffold dependence. Agentic and SWE-bench results depend heavily on the harness, tool set, and token budget. Your stack may not reproduce them.
• Anecdotal tests. A hands-on review ran the model through llama.cpp using a GGUF build and had it fix an inheritance-logic bug in a multi-container Docker app, which it did. The same reviewer found its reasoning verbose on an image task and its translations weak in low-resource languages. One reviewer's tests are not a benchmark, but they match the "strong coder, imperfect generalist" picture.

Is Ornith-1.5-122B worth testing?

If you want an MIT-licensed open-weight model for coding agents and tool use that fits on a single high-memory machine, the card makes it a reasonable candidate. The strongest case is software-engineering work, where it leads comparable-size models by wide margins. If your workload is deep web research or maximal reasoning with tools, the DeepSeek-V4-Flash numbers suggest a larger model may still score higher.

A sensible approach is to run your own repository tasks through the same agent harness for Ornith and one or two rivals, and compare pass rates and token use. The "leaner outputs" claim is easy to measure that way.