ornith-1.5-9b on Tenstorrent p150 and p150x2 and p150x4

Serves ornith-ai/Ornith-1.5-9B by Ornith AI at revision 489cb979. See the original model card for training and evaluation details.

Ornith-1.5-9B, a 9B-parameter dense reasoning and coding model, served with vLLM with tool calling and separated reasoning output.

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

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

At a glance

Hardware	p150, p150x2, p150x4
Context	262,144 tokens
License	mit

Intended use

Agentic coding, reasoning and tool calling through an OpenAI-compatible API.

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

Quickstart

tt serve tt-hous/ornith-1.5-9b

tt serve downloads the Docker image and the ornith-ai/Ornith-1.5-9B at 489cb97981b8654bcfcf30ce1f94ed1b62e07b53 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 Application startup complete.

Without tt-cli: tt-model serve tt-hous/ornith-1.5-9b.

Serving profiles

Select a different serving profile using --profile.

profile	hardware	mesh	max_num_seqs	max_model_len	description
p150	p150	P150	32	262144	One Blackhole ASIC. Full 262144-token context; lowest hardware bar.
p150x2	p150x2	P150x2	32	262144	Two Blackhole ASICs (one p300c board). Full 262144-token context.
p150x4 (default)	p150x4	P150x4	32	262144	Four Blackhole ASICs (two p300c boards).

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-1.5-9B" — 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, apart from content.

The model card's sampling recipe for general tasks is temperature 1.0, top_p 0.95, top_k 20, presence penalty 1.5 (the recipe the GPQA Diamond run below used); for precise coding tasks it is temperature 0.6, top_p 0.95, top_k 20.

Expected performance

eval	this package	reference	prompts	notes
GPQA Diamond	78.3	86.4	198	mean of 2 runs here (seeds 42 / 43: 75.3 / 81.3); the reference is a mean of 5 runs

GPQA Diamond ran on the p150 profile with the model card's recipe: temperature 1.0, top_p 0.95, top_k 20, presence penalty 1.5, and a 32,768-token output budget. The reference is the Ornith AI model card's figure.

profile	ISL	OSL	concurrency	N	TTFT (ms)	prefill tok/s/u	decode tok/s/u	total output tok/s	E2EL (s)
p150	128	128	1	1	97	1,313	6.0	6	21.35
p150	128	128	8	1	882	145	5.9	46	22.27
p150	128	128	32	1	3,785	34	5.9	162	25.29
p150	1,024	256	1	1	196	5,224	6.0	6	43.03
p150	1,024	256	8	1	1,629	629	5.9	46	44.64
p150	1,024	256	32	1	6,864	149	5.9	163	50.29
p150	4,096	256	1	1	558	7,343	5.9	6	43.84
p150	4,096	256	8	1	4,508	909	5.9	43	47.98
p150	4,096	256	32	1	18,376	223	5.7	130	62.99
p150	16,384	256	1	1	2,442	6,709	5.7	5	47.55
p150	16,384	256	8	1	19,552	838	5.6	32	64.89
p150	16,384	256	32	1	82,547	198	4.9	41	134.61
p150	32,768	256	1	1	5,687	5,762	5.4	5	53.22
p150	32,768	256	8	1	45,506	720	5.3	22	93.32
p150	32,768	256	32	1	185,051	177	5.3	22	232.73
p150	65,536	256	1	1	14,704	4,457	4.9	4	67.08
p150	65,536	256	8	1	114,339	573	4.9	9	166.76
p150	131,072	256	1	1	42,781	3,064	4.1	2	104.86
p150x2	128	128	1	1	68	1,886	9.7	10	13.23
p150x2	128	128	8	1	576	222	9.6	74	13.79
p150x2	128	128	32	1	2,223	58	9.5	264	15.53
p150x2	1,024	256	1	1	145	7,068	9.6	10	26.65
p150x2	1,024	256	8	1	1,150	890	9.6	74	27.72
p150x2	1,024	256	32	1	4,623	222	9.5	260	31.47
p150x2	4,096	256	1	1	429	9,539	9.5	9	27.16
p150x2	4,096	256	8	1	3,426	1,196	9.5	68	30.24
p150x2	4,096	256	32	1	13,745	298	9.3	199	41.21
p150x2	16,384	256	1	1	1,838	8,913	9.2	9	29.48
p150x2	16,384	256	8	1	14,698	1,115	9.2	48	42.42
p150x2	16,384	256	32	1	58,683	279	7.8	60	91.27
p150x2	32,768	256	1	1	4,205	7,792	8.8	8	33.06
p150x2	32,768	256	8	1	33,599	975	8.8	33	62.54
p150x2	32,768	256	32	1	127,117	258	8.8	33	156.00
p150x2	65,536	256	1	1	10,548	6,213	8.2	6	41.81
p150x2	65,536	256	8	1	78,830	831	8.2	14	110.05
p150x2	131,072	256	1	1	29,784	4,401	7.1	4	65.88
p150x2	131,072	256	8	1	202,876	646	7.1	5	238.91
p150x4	128	128	1	1	66	1,948	15.4	15	8.32
p150x4	128	128	8	1	769	166	17.2	126	8.14
p150x4	128	128	32	1	2,295	56	15.3	386	10.62
p150x4	1,024	256	1	1	97	10,597	15.3	15	16.71
p150x4	1,024	256	8	1	742	1,380	17.2	132	15.54
p150x4	1,024	256	32	1	2,952	347	15.2	415	19.73
p150x4	4,096	256	1	1	258	15,907	15.2	15	16.98
p150x4	4,096	256	8	1	2,059	1,989	17.1	121	16.97
p150x4	4,096	256	32	1	8,173	501	14.9	324	25.24
p150x4	16,384	256	1	1	1,052	15,573	14.8	14	18.23
p150x4	16,384	256	8	1	8,420	1,946	16.6	86	23.76
p150x4	16,384	256	32	1	34,141	480	12.6	101	54.46
p150x4	32,768	256	1	1	2,342	13,993	14.4	13	20.10
p150x4	32,768	256	8	1	18,546	1,767	16.0	59	34.49
p150x4	32,768	256	32	1	70,032	468	16.1	60	85.92
p150x4	65,536	256	1	1	5,585	11,735	13.4	10	24.55
p150x4	65,536	256	8	1	42,898	1,528	13.5	25	61.83
p150x4	65,536	256	32	1	166,320	394	13.5	25	185.21
p150x4	131,072	256	1	1	14,983	8,748	11.9	7	36.35
p150x4	131,072	256	8	1	106,942	1,226	12.0	10	128.25

Means over one run per row on each profile's mesh, with a closed-loop latency client: random-token prompts of exactly ISL tokens, output pinned to OSL tokens with ignore_eos, greedy, streamed; three warm-up requests per input length; each run sends 4 requests per user up to 4,096 input tokens, 2 per user at 16,384 and 1 per user above. TTFT is time to the first streamed token; decode tok/s/u is 1 / TPOT with TPOT = (E2EL - TTFT) / (OSL - 1); prefill tok/s/u is ISL / TTFT, so above concurrency 1 it includes queueing time. The rows are selected from a sweep at 1 to 32 concurrent users; rows predicted to exceed 300 s end to end were skipped. Peak aggregate output across the sweep: 163 tok/s on p150 (1,024 in, 256 out, 32 users), 264 on p150x2 (128 in, 128 out, 32 users) and 415 on p150x4 (1,024 in, 256 out, 32 users).

Limitations

• Measured on one Blackhole chip (p150), two (p150x2) and four (p150x4). GPQA Diamond is the only accuracy eval and ran on p150 only. No agentic-coding benchmark (SWE-bench, Terminal-Bench) was run.
• Not for high-throughput serving on p150: decode is about 6 tokens per second per user at every concurrency.
• With 32 concurrent requests whose generations together exceed the 262,144-token pool, the engine exits (no free device state slot). Keep concurrent long generations below the pool, for example --max-num-seqs 16 on p150.
• On p150 and p150x2 a logprobs request is rejected with HTTP 400; only p150x4 returns logprobs.
• min_p, bad_words, logit_bias, allowed_token_ids, min_tokens, structured outputs and top_logprobs above 0 are rejected with HTTP 400 on every profile.

Feedback

Questions or problems with this package: open a discussion at https://huggingface.co/tt-hous/ornith-1.5-9b/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: support@tenstorrent.com.

Provenance

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

component	built from
tt-metal	c0976f241fa369f0a12a561141f0ffc91f7ac087
vLLM	v0.26.0
vllm-tt-plugin	9bab7d527d4adbece91de7338302c99e652fd642
code/ digest	13978ba8a38bed56 (sha256, first 16 hex digits)
built	2026-10-23T19:47:44+00:00 by tt-model 0.1.0