AI Tools Radar

What Is Ornith-1.0, the New Open-Source AI Coding Agent?

Published 2026-06-30 · Researched and price-checked on official sources

Ornith-1.0 is a new open-source AI coding agent from DeepReinforce that autonomously writes, debugs, and refactors code through your terminal. On standard coding benchmarks it matches Claude Opus 4.7, and it is MIT-licensed and free. The catch: you need your own GPU hardware to run it, and the full 397B model demands a multi-GPU setup.

Key facts

What matters

Ornith-1.0 is MIT-licensed and free: no subscription, no per-token billing. The LICENSE file is confirmed on the GitHub repository.
The 397B variant scores 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, matching Claude Opus 4.7 (70.3 and 80.8 respectively), per the official README benchmarks.
Running it is not free: the 9B model needs an 80GB GPU, and the 397B MoE model requires a multi-GPU node with tensor parallelism.
It ships as an OpenAI-compatible server via vLLM, so any tool that talks to an OpenAI endpoint can use it as a drop-in replacement.
Available in four sizes: 9B Dense, 31B Dense, 35B MoE, and 397B MoE, all post-trained on Gemma 4 and Qwen 3.5 base models.

What exactly is Ornith-1.0?

Ornith-1.0 is a family of open-source large language models built specifically for agentic coding: writing, debugging, and refactoring code autonomously through a terminal. It was released in June 2026 by DeepReinforce, a research group focused on reinforcement learning for code generation.

The models are available on HuggingFace and GitHub under the MIT license (verified as of June 30, 2026). That means you can download the weights, fine-tune them, and deploy them commercially with no restrictions. The project had 525 GitHub stars and 50 forks as of this writing, and the 397B HuggingFace model page shows 1,622 downloads and 168 likes.

Ornith-1.0 comes in four sizes, each post-trained on top of Gemma 4 and Qwen 3.5 foundation models:

Ornith-1.0-9B Dense: fits on a single 80GB GPU for edge deployment
Ornith-1.0-31B Dense: mid-range option for single-GPU or small-node setups
Ornith-1.0-35B MoE: mixture-of-experts with lower active parameters per token
Ornith-1.0-397B MoE: frontier-scale model requiring multi-GPU tensor parallelism

All variants support a 256K token context window (262,144 tokens), which is enough to ingest an entire medium-sized codebase in one prompt. The models are published in multiple precision variants including BF16 and FP8, with the FP8 checkpoints optimized for memory-efficient serving.

What makes Ornith-1.0 different: self-scaffolding

The core innovation behind Ornith-1.0 is a training method DeepReinforce calls self-scaffolding. Most AI coding models are trained with a fixed harness: a human-written wrapper that tells the model how to approach a coding task, run tests, and iterate. The harness is static, so the model never learns to improve how it attacks problems.

Ornith-1.0 flips this. During reinforcement learning, the model learns to generate both the solution code and the scaffold that drives the solution process. By jointly optimizing the scaffold and the resulting code, the model discovers better search trajectories on its own. It figures out which problem-solving strategies work and which do not, without a human scripting that discovery process.

In plain terms: Ornith-1.0 does not just get better at writing correct code. It gets better at figuring out how to write correct code. That is a deeper capability, and the benchmark numbers bear it out: the 397B model scores 77.5 on Terminal-Bench 2.1 (Terminus-2) and 82.4 on SWE-Bench Verified, per the official README. For context, Claude Opus 4.7 scores 70.3 and 80.8 on those same benchmarks, and DeepSeek V4 Pro (1.6T parameters) scores 64 and 80.6.

Who should use Ornith-1.0 (and who should skip it)

Ornith-1.0 is not for everyone. It is an open-weight model you host yourself. There is no managed API, no web dashboard, no SaaS plan. That self-hosting requirement defines who it works for.

Good fit if you:

Already own or rent GPU hardware (an 80GB A100 or H100 for the 9B, or a multi-GPU node for the 397B)
Want an AI coding agent that can run entirely on your infrastructure with no data leaving your network
Need an OpenAI-compatible endpoint you can plug into existing tools via vLLM
Are comfortable with model deployment, quantization, and inference server setup
Want to fine-tune a coding model on your private codebase under a permissive MIT license

Skip it if you:

Do not own a GPU with at least 80GB of VRAM. Cloud GPU rental adds monthly cost that may exceed a Copilot subscription
Want a zero-setup experience where you install an IDE plugin and start coding
Are an individual developer with a laptop. Even the 9B model will not run on consumer hardware
Need guaranteed uptime and support. There is no SLA, no paid tier, no vendor to call

Ornith-1.0 vs GitHub Copilot, Cursor, and Claude Code

The table below puts Ornith-1.0 side by side with the three most popular paid AI coding tools. Ornith-1.0 is the only option that is free and self-hosted, but that freedom comes with infrastructure responsibility.

The hardware column is the honest differentiator: GitHub Copilot runs on Microsoft's servers and costs $10 to $39 per month. Ornith-1.0 costs nothing for the software, but renting an 80GB A100 on cloud providers like Lambda Labs or RunPod runs roughly $1 to $2 per hour. If you use it for more than 10 to 20 hours per month, a Copilot subscription is cheaper than the GPU rental alone. That is before you account for setup time.

The sweet spot for Ornith-1.0 is teams that already own GPU capacity and want a coding agent that never phones home. For everyone else, the paid tools remain the more practical choice today.

How to run Ornith-1.0

The README provides vLLM deployment recipes. You pull the weights from HuggingFace, launch an OpenAI-compatible server, and point any tool that speaks the OpenAI API at your local endpoint. The shared alias is Ornith-1.0, and you set the MODEL environment variable to the checkpoint you want.

Key parameters from the official setup:

9B Dense: fits on one 80GB GPU (A100 or H100)
397B MoE: requires tensor parallelism across multiple GPUs
Context: 256K tokens (262,144)
Precision options: BF16 for accuracy, FP8 for memory efficiency
Supported frameworks: vLLM with reasoning_content key alignment via Harbor

The evaluation harness used in the official benchmarks (Harbor/Terminus-2) runs with 32 CPU cores, 48GB RAM, and a 4-hour timeout per task. That is not the minimum to run the model, but it gives a sense of what the model was tested on.

At a glance

Feature	Ornith-1.0	GitHub Copilot	Cursor	Claude Code
Software cost	Free (MIT)	$10–$39/mo	$20–$40/mo	API per-token
Hardware needed	Your own GPU(s)	None (cloud)	None (cloud)	None (cloud)
Largest model	397B MoE	Proprietary	Proprietary	Proprietary
Context window	256K tokens	Varies by model	Varies by model	200K tokens
Self-hosting	Yes	No	No	No
License	MIT	Proprietary	Proprietary	Proprietary
IDE integration	Any OpenAI-compatible client	VS Code, JetBrains, Neovim	Standalone IDE	Terminal, VS Code

Ornith-1.0 vs paid AI coding tools at a glance (costs verified June 30, 2026)

FAQ

Is Ornith-1.0 actually free?

The software is free under the MIT license. No subscription, no per-token cost, no usage caps. But you need GPU hardware to run it. The 9B model requires an 80GB GPU (A100 or H100), and the 397B model needs multiple GPUs with tensor parallelism. If you rent cloud GPUs, the hourly cost can exceed a paid tool subscription depending on usage.

How does Ornith-1.0 compare to GitHub Copilot?

Ornith-1.0 is an autonomous terminal-based coding agent that plans, writes, and debugs code end-to-end. GitHub Copilot is an inline autocomplete and chat assistant inside your IDE. Copilot is easier to set up but less autonomous. Ornith-1.0 handles larger, multi-file tasks on its own but requires GPU infrastructure and manual deployment. On benchmarks, the 397B Ornith model outperforms the coding ability of models Copilot currently uses. Ease of use strongly favors Copilot for most developers.

Can I run Ornith-1.0 on my laptop?

No. Even the smallest 9B Dense model requires an 80GB GPU like an NVIDIA A100 or H100. These are data-center GPUs, not consumer hardware. Consumer GPUs with 24GB of VRAM (like an RTX 4090) cannot load even the smallest variant at full precision. There is no quantized consumer-friendly release as of June 2026.