Fine-tuning Ornith 9B on one H200 GPU: $85, 25 hours

Key Takeaways

• Fine-tuning Ornith-1.0-9B on 61,000 examples cost $84.88 on a single H200 GPU Droplet at $3.44/hour
• The fine-tune adds structured reasoning summaries without degrading inference latency or throughput
• The 9B model fits in bf16 on a single 80GB GPU using roughly 19GB of VRAM

DigitalOcean published a full walkthrough of fine-tuning Ornith-1.0-9B, an open-source agentic coding model from DeepReinforce AI, on a single H200 GPU Droplet. The total bill: $84.88 for 24.7 hours of compute at $3.44/hour. The fine-tune trained on 61,000 reasoning-summarization examples using LLaMA-Factory, and the tutorial includes serving benchmarks via vLLM to show what happens to latency and throughput after the modification.

This matters because the fine-tune doesn't touch the model's coding performance or agentic scaffolding behavior. It only changes how reasoning is surfaced to end users. Instead of dumping raw chain-of-thought or hiding it entirely, the modified model produces a structured summary: title, subtitle, plain-language summary, and a current-task field. The question the tutorial answers is whether adding this output behavior costs anything at inference time. According to their benchmarks, it doesn't.

What is Ornith-1.0-9B and why does it matter?

Ornith-1.0-9B is the smallest model in DeepReinforce AI's Ornith-1.0 family, which includes 31B dense, 35B MoE, and 397B MoE checkpoints. It's MIT-licensed, built on top of Qwen 3.5, and trained specifically for agentic coding tasks. The architectural differentiator isn't size but training method: instead of pairing the model with a fixed agent harness, Ornith uses reinforcement learning to construct its own scaffold alongside solutions, optimizing both jointly.

DeepReinforce reports the 9B variant hits 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified. Those numbers match or beat larger open models on comparable benchmarks. The model fits in bf16 on a single 80GB GPU, requiring about 19GB of VRAM, which explains why a single H200 handles both training and serving without multi-GPU coordination.

Training setup and actual costs

The tutorial uses LLaMA-Factory running on a DigitalOcean H200 GPU Droplet. The dataset, SupraLabs/reasoning-summaries-61k, contains 61,000 examples pairing raw reasoning traces with structured summaries. Each example includes title, subtitle, summary, and current-task fields. The data lineage is traceable to named, licensed upstream sources rather than scraped or proprietary material.

At $3.44/hour, the H200 Droplet pricing sits in the typical cloud GPU range. For comparison, Lambda Labs charges around $2-3/hour for H100s, and RunPod's H100 pricing hovers around $2.50/hour depending on availability. The H200's 141GB HBM3e memory and 4.8 TB/s bandwidth give it headroom that the H100 lacks for larger models, though that advantage is less relevant for a 9B parameter model that fits comfortably on 80GB hardware.

Serving benchmarks: latency and throughput

The tutorial deploys the fine-tuned model using vLLM on the same H200 Droplet and tests time-to-first-token (TTFT) and throughput at 1, 5, and 20 concurrent requests. The key finding: the fine-tune doesn't degrade inference performance compared to the base model. Output format changes, when they don't alter the model's underlying task behavior, are often assumed to be free at serving time. This tutorial provides actual numbers rather than assumptions.

The distinction matters for production deployments. Model providers building user-facing agentic products typically choose between two bad defaults. Hide reasoning entirely, which makes debugging and trust-building harder. Or dump raw chain-of-thought on users, which is verbose, repetitive, and sometimes surfaces information you'd rather not expose. A model that natively produces clean structured summaries skips the need for a separate summarization pass or a second model call.

When this approach makes sense

This fine-tuning pattern fits a specific use case: you want to change how a model presents its reasoning without changing what it can do. If your goal is improved accuracy, better tool-use, or stronger agentic behavior, you need a different kind of fine-tune and a different evaluation setup. The tutorial is explicit about this boundary.

The $85 price point puts this within reach of individual developers and small teams. You don't need a multi-GPU cluster. You don't need distributed training expertise. LLaMA-Factory handles the training loop, and vLLM handles serving. The entire workflow runs on a single GPU Droplet that you can spin up on demand.

Reproducibility and data lineage

The tutorial emphasizes that the training data comes from traceable, licensed sources. This matters for anyone considering whether to reproduce the experiment. Proprietary or scraped datasets create legal uncertainty. The SupraLabs/reasoning-summaries-61k dataset's documented lineage means you can evaluate the license terms before committing compute.

Ornith-1.0-9B's MIT license similarly removes friction. You can modify and deploy the fine-tuned model commercially without negotiating additional terms. The combination of open model, open data, and documented toolchain makes this a reproducible reference implementation rather than a one-off demo.

Frequently Asked Questions

How much VRAM does Ornith-1.0-9B require for fine-tuning?

The model requires approximately 19GB of VRAM in bf16 precision, fitting comfortably on a single 80GB GPU like the H200.

Does fine-tuning for structured reasoning summaries hurt inference speed?

According to the benchmarks in this tutorial, no. The fine-tuned model showed no degradation in time-to-first-token or throughput compared to the base model.

Can I run this fine-tune on an H100 instead of an H200?

Yes. The 9B model's memory requirements are well within H100's 80GB capacity. The H200's extra headroom is more relevant for larger models in the Ornith family.

What benchmarks does Ornith-1.0-9B achieve on coding tasks?

DeepReinforce reports 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified, matching or exceeding larger open models.

Is the training data commercially usable?

The SupraLabs/reasoning-summaries-61k dataset has documented lineage from named, licensed upstream sources. Check the specific license terms before commercial deployment.