What happened
Ornith describes Ornith-1.5 as an extension of Ornith-1.0, which it says was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training and post-training. The new release expands what the company calls its self-improvement loop. Instead of relying only on a fixed set of human-curated tasks and manually designed harnesses, the system generates training tasks, searches for strategies to solve them, and uses reinforcement learning to improve the policy. Ornith says the process jointly optimizes task generation, scaffold construction and solution rollouts. Its model card points readers to a blog for more detail on task, harness and rollout reward design.
The card calls Ornith-1.5-122B the most capable and efficient member of its family. It is a 122B MoE model, with a quantized Ornith-1.5-122B-Mobile variant described as edge-deployable on mobile devices. The model is available on Hugging Face under an MIT license. The page includes examples for loading it with Transformers and serving it with vLLM or SGLang through OpenAI-compatible chat-completions endpoints. It specifies recent runtime versions: Transformers 5.8.1 or later, vLLM 0.19.1 or later, and SGLang 0.5.9 or later. Its examples also show image-and-text input, consistent with the repository’s image-text-to-text tags.
Ornith’s model card reports results across coding, reasoning and agentic benchmarks, comparing the model with several other large models. Reported scores include 84.5 on SWE-bench Verified, 76.9 on SWE-bench Pro, 75.2 on WideSearch and 74.6 on MCP-Atlas. The card describes different evaluation conditions by benchmark: SWE-bench uses the OpenHands harness with a 256K context window and anti-hacking safeguards; MCP-Atlas uses a 500-task public subset, thinking mode and a ten-minute timeout per task. The listed comparisons do not show Ornith leading every benchmark: for example, its reported Terminal-Bench 2.1 score is 81.2, below a listed competitor’s 83.4.
The release notes also warn that this is a reasoning model whose assistant turn opens with a <think> block. The serving recipes configure reasoning and tool-call parsers so reasoning content and tool calls can be returned in separate fields. The page provides installation and serving examples, but it does not make a general-availability or hosted-inference claim: the repository lists no inference providers for the model.
What the archive adds
New Runtime covered Ornith-1.5-122B on 2026-11-07 in an earlier note about the 122B MoE model and its MIT license. This is a new fact on a known story: the current model card describes the expanded self-improvement loop, serving requirements and benchmark methodology. An earlier New Runtime note on 2026-10-27 discussed a paper on self-improvement in language models, a related but distinct approach.
Trend signal
This release is relevant to harness architecture outlives model choice, but it is not strong confirmation of the pattern. Ornith’s account makes task generation, scaffolds, rollouts and evaluation harnesses part of the training story rather than treating the model weights as the only relevant component. The release does not establish how durable or transferable those choices are across models or workloads.
Why it matters
For teams evaluating open models, the model name and parameter count are not enough to predict whether a model will work in an agent stack. Ornith’s card makes clear that inference support, chat-template behavior, parsers and harness-specific settings are part of the deployment decision. A model that can call tools through an OpenAI-compatible endpoint may still require recent serving runtimes and configuration to expose those calls properly.
The self-improvement claim also points to a useful distinction in model evaluation: benchmark performance is tied to how tasks are generated, what scaffolds are used and how rewards are assigned. Teams should treat published scores as evidence about the named test setup, not as a guarantee of performance on their own production tasks. The reported results can help identify candidate areas for testing, while local task suites and checks remain necessary before adoption.
The model card’s stated mobile variant could matter for teams exploring local or edge deployment, but the card’s claim alone is not a measurement of practical mobile performance. Hardware, quantization, latency and quality trade-offs determine whether that option is useful in a real application. Teams considering the model should begin with compatibility and workload fit rather than assuming that the MoE designation or mobile variant makes it inexpensive to operate.
Limits and uncertainty
The training-loop description is Ornith’s account of its own process; the model card does not independently establish which part of the loop caused particular benchmark results. The scores are reported by the model publisher, and the cited evaluations use differing harnesses, context limits, judges and task subsets, so cross-benchmark comparisons should be treated cautiously. The card’s statements about efficiency and mobile edge deployment are not accompanied here by a complete hardware, latency or cost comparison. The practical behavior of the model, its quantized variant and the required serving stack should be tested independently.
Reader takeaway
Ornith-1.5-122B is more than a model-weight release: its card describes a training loop that generates tasks and improves scaffolds alongside solution rollouts, and it makes deployment requirements explicit. Treat its benchmark table as a publisher-reported starting point, then test the model with your own tools, harness and workload before drawing conclusions.