Ornith-1.5

Learning Objectives

• Understand what makes Ornith-1.5's license genuinely different from the other large open-weight releases of 2026
• Evaluate whether the flagship's benchmark parity with a closed frontier model translates to your own workload
• Judge the hardware and operational cost of self-hosting each of the three sizes

What Is Ornith-1.5?

Ornith-1.5 is a family of open-weight language models released on August 19, 2026 by DeepReinforce, an AI research lab previously known for CUDA-L1 and the IterX code-agent optimization loop. It comes in three sizes: a 397 billion parameter mixture-of-experts (MoE) flagship, a 35 billion parameter MoE model that activates roughly 3 billion parameters per token, and a 9 billion parameter dense model with a quantized build small enough for phones.

Two things make it worth a page of its own. The first is the license. The second is how it was trained.

The License Is the Headline

Ornith-1.5 ships under a plain MIT license — the actual MIT text, not a vendor license wearing MIT's name. There is no revenue threshold, no monthly-active-user ceiling, no attribution requirement, no clause forbidding you from training competing models, and no geographic carve-out.

That last point matters more than it sounds. 2026 has been the year "open weights" quietly stopped meaning "open license." Kimi K3's weights landed under custom terms requiring a separate agreement to host inference above $20 million in revenue. Qwen3.8-Max carried a $50 million trigger. MiniMax H3 excluded the European Union, the United Kingdom, South Korea and the United States outright — meaning a US developer could not legally use the weights at all. Against that backdrop, a frontier-scale model under unmodified MIT is the outlier, not the norm.

Self-Scaffolding Reinforcement Learning

Most reinforcement-learning pipelines for coding models train against a fixed set of human-written tasks inside human-designed harnesses. Ornith inverts that. During training the model proposes its own tasks, builds a scaffold for approaching each one, and generates the solution rollouts that become the learning signal. The reward optimizes the strategy and the solution together, so task-specific approaches emerge without an engineer designing them in advance.

DeepReinforce reports tuning the task-proposal reward around validity, novelty, and difficulty, targeting roughly a 0.2 success rate — hard enough to be informative, easy enough to be learnable. The practical claim is that a model which writes its own curriculum is not bounded by the imagination of the people who would otherwise have written it.

Treat that as a promising result rather than a settled one. It is a single lab's reported method on a single family, and the interesting comparison — whether self-scaffolding still helps at the next scale up — has not happened yet.

Benchmarks

The shape of these numbers is more useful than any single one. On agentic terminal work the flagship edges ahead of a closed frontier model; on general reasoning it sits just behind; on the harder software-engineering benchmark it trails by three points. That is a model that has been optimized hard for agentic coding and is competitive rather than dominant elsewhere — which is exactly what its training method would predict.

The smaller sizes land where you would expect: the 35 billion parameter MoE scores 68.5 on Terminal-Bench 2.1 and 79.0 on SWE-Bench Verified, and the 9 billion parameter dense model scores 47.0 and 70.6.

Pricing

There is no vendor pricing to compare because there is no vendor product to buy. The cost of Ornith-1.5 is entirely the cost of the hardware you run it on, which is the trade the MIT license buys you: no per-token bill and no vendor relationship, in exchange for owning the operational problem yourself.

What It Takes to Run

The flagship is roughly 800 gigabytes at bf16 precision and needs eight-way tensor parallelism — on the order of eight H200 cards on a single node. That is a serious infrastructure commitment, not a weekend experiment, and it is the practical reason most readers will never run the 397 billion parameter model despite the license permitting it.

The 35 billion parameter MoE is the interesting middle: because it activates only about 3 billion parameters per token, it serves far more cheaply than its total size suggests. The 9 billion parameter dense model, including its quantized mobile build, is the one that runs on ordinary hardware.

Strengths

• Genuinely permissive license — unmodified MIT with no revenue trigger, no user ceiling, and no geographic exclusion, which is rare at this scale in 2026
• Frontier-adjacent agentic coding — edges a closed frontier flagship on Terminal-Bench 2.1, the benchmark closest to real terminal-agent work
• Three sizes with a real spread — a 9 billion parameter dense model for local use through a 397 billion parameter flagship, rather than one size and a token small variant
• Long context — 262,144 tokens natively, extendable toward one million with YaRN scaling
• No vendor lock-in — no API to be deprecated, no pricing to change, no terms that can be revised under you

Limitations and Considerations

• No hosted API — there is no first-party endpoint and no price list, so evaluating the model at all requires standing up inference yourself
• Flagship hardware cost is prohibitive for individuals — roughly eight H200 cards for the 397 billion parameter model
• Not dominant outside agentic coding — trails Claude Opus 4.8 on GPQA Diamond and DeepSWE; the specialization is real and cuts both ways
• Single-lab benchmark reporting — the numbers are DeepReinforce's own, averaged over five runs, and have not been independently reproduced
• Thin public track record — the lab is young and the family is two releases old, so there is little history on which to judge support, security response, or release cadence

Key Takeaways

• Ornith-1.5 is a three-size open-weight family from DeepReinforce, released August 19, 2026 under an unmodified MIT license with no revenue, user, or geographic restrictions
• The 397 billion parameter flagship scores 86.1 on Terminal-Bench 2.1 against Claude Opus 4.8's 85.0, and 86 on SWE-Bench Verified — frontier-adjacent on agentic coding, slightly behind on general reasoning
• Its training method is self-scaffolding reinforcement learning: the model proposes its own tasks and builds its own scaffolds rather than learning inside human-designed harnesses
• The license is the story. In a year when Kimi K3, Qwen3.8-Max and MiniMax H3 all shipped "open weights" under restrictive custom terms, plain MIT at this scale is the exception
• There is no hosted API, so the real cost is hardware — about eight H200 cards for the flagship, while the 9 billion parameter dense model is the one most people can actually run