ornith-1.0-35b

ornith-1.0-35b on Tenstorrent p150 and p150x2 and p150x4 and p100

Serves ornith-ai/Ornith-1.0-35B by Ornith AI at revision 5df2ed3f. See the original model card for training and evaluation details.

Ornith-1.0-35B, a 35B-parameter mixture-of-experts model for agentic reasoning and coding, served with vLLM with tool calling and separated reasoning output.

Runs on Tenstorrent p150 or p150x2 or p150x4 or p100 — see the serving profiles below.

Packaged and published with tt-model-manager 0.1.0 (manifest schema 5.1).

At a glance

Intended use

Agentic coding and reasoning through an OpenAI-compatible API, for example from OpenCode.

Prerequisites

• The Tenstorrent CLI, tt: uv tool install tenstorrent (or use tt-model alone, see the Quickstart)
• Docker
• Tenstorrent hardware: p150 or p150x2 or p150x4 or p100

Quickstart

tt serve tt-hous/ornith-1.0-35b

tt serve downloads the Docker image and the ornith-ai/Ornith-1.0-35B at 5df2ed3f675c7beaa490328cc70bb573b65fb660 weights (into your HF cache; they are not in the image), then starts an OpenAI-compatible server on port 20000 (or the next free port, if that one is busy). The first start compiles kernels for your device, which takes several minutes; the server is ready when it logs Server ready after.

Without tt-cli: tt-model serve tt-hous/ornith-1.0-35b.

To use it from OpenCode, add an OpenAI-compatible provider with the base URL the serve command printed and the model ornith-ai/Ornith-1.0-35B.

Serving profiles

Select a different serving profile using --profile.

Capabilities

The server speaks the OpenAI API at http://127.0.0.1:20000/v1 (or whichever port your serve command reported; chat completions, completions, and /v1/models). Pass "model": "ornith-ai/Ornith-1.0-35B" — the weights id, not this package's name.

Tool calling is enabled (--tool-call-parser qwen3_xml): a request that passes tools comes back with tool_calls and finish_reason: tool_calls.

Reasoning output is separated (--reasoning_parser qwen3): the thinking text arrives in reasoning (reasoning_content on vLLM before 0.11.1), apart from content.

The model card's sampling for general tasks is temperature 0.6, top_p 0.95, top_k 20.

Expected performance

Evals ran on 2026 Sep 30 on p150x4, against this package's image, with lm-evaluation-harness through the chat API: temperature 0.6, top_p 0.95, top_k 20, thinking off, 0-shot, one request at a time. The Ornith AI model card publishes no GPQA Diamond, LiveCodeBench or IFEval figure (it reports Terminal-Bench, SWE-bench, NL2Repo and ClawEval), so the reference column is empty.

Medians of three sequential requests per row, measured with vllm bench serve on one QuietBox 2 (two P300C boards): one ASIC for p150, both ASICs of one board for p150x2, all four for p150x4. Random prompts of exactly ISL tokens, output pinned to OSL tokens, greedy. Prefill tok/s/u is ISL divided by TTFT, decode tok/s/u is 1,000 divided by TPOT (ms), and E2EL is TTFT plus 127 times TPOT. The last p150x4 row fills the full 262,144-token context. No p100 row: that profile has not been measured.

Limitations

• One request at a time on every profile (max_num_seqs 1).
• Tested on one QuietBox 2 only: p150 on one of its chips, p150x2 on one of its boards, p150x4 on both. The p100 profile has not been run on a p100 board; it has no speed or eval data.
• Accuracy was measured on subsets with thinking off, on p150x4 only. With a 4,096-token cap, 13 of 40 GPQA Diamond and 22 of 50 LiveCodeBench responses ran out of budget before answering. No agentic-coding benchmark was run.
• Not for interactive use on p150: decode is about 11.5 tokens per second, and a 32,768-token prompt takes about 43 s to first token.

Feedback

Questions or problems with this package: open a discussion at https://huggingface.co/tt-hous/ornith-1.0-35b/discussions — that is what reaches its author. A problem with the tt tooling itself: tt report issue (collects your environment and opens a prefilled issue against tenstorrent/tt-cli). Product feedback: [email protected].

Provenance

The exact sources the image was built from — code/ in this repo is byte-identical to the model code inside the image: