laguna-xs-2.1 on Tenstorrent p150x2 and p150x4

Serves poolside/Laguna-XS-2.1 by poolside at revision c5f36269. See the original model card for training and evaluation details.

Laguna-XS-2.1, poolside's 33B-parameter mixture-of-experts coding model (3B active parameters), served with vLLM with tool calling and separated reasoning output.

Runs on Tenstorrent 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	p150x2, p150x4
Context	131,072 tokens
License	openmdw-1.1

Intended use

Agentic coding 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: p150x2 or p150x4
• p150x2: two P150 boards or one P300C board. p150x4: four P150 boards or two P300C boards (a QuietBox 2).

Quickstart

tt serve tt-hous/laguna-xs-2.1

tt serve downloads the Docker image and the poolside/Laguna-XS-2.1 at c5f36269bbdbd3f27fddc9a9f9dbae0cf2cf57db 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/laguna-xs-2.1.

Serving profiles

Select a different serving profile using --profile.

profile	hardware	mesh	max_num_seqs	max_model_len	description
p150x2 (default)	p150x2	P150x2	1	131072	Two P150 ASICs, with ring/two-link fabric and canonical prefix caching.
p150x4	p150x4	P150x4	8	131072	Four-P150 regression profile with prefix caching explicitly disabled.

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": "poolside/Laguna-XS-2.1" — the weights id, not this package's name.

Tool calling is enabled (--tool-call-parser poolside_v1): a request that passes tools comes back with tool_calls and finish_reason: tool_calls.

Reasoning output is separated (--reasoning_parser poolside_v1): the thinking text arrives in reasoning, apart from content.

Expected performance

eval	this package	reference	prompts	notes
GPQA Diamond, exact match, thinking off	42.5%	—	40 of 198	standard error 7.9
IFEval, prompt-level strict / loose, thinking off	73.0% / 76.0%	—	100 of 541	standard error 4.5 (strict)
IFEval, instruction-level strict, thinking off	81.6%	—	100 of 541

Evals ran on 2026 Sep 30 on p150x2, against this package's image, with lm-evaluation-harness through the chat API: the model card's sampling (temperature 1.0, top_p 1.0, top_k 20), thinking off, 0-shot, one request at a time, 4,096 output tokens for GPQA and 2,048 for IFEval. The original model card reports no figure for either benchmark. It reports SWE-bench Verified 70.9%, SWE-bench Multilingual 63.1%, SWE-Bench Pro (public dataset) 47.6% and Terminal-Bench 2.0 37.5%, measured with thinking on; none of these has been run against this package.

profile	ISL	OSL	concurrency	N	TTFT (ms)	prefill tok/s/u	decode tok/s/u	E2EL (s)
p150x2	128	512	1	1	215	595	20.0	25.80
p150x2	1,024	512	1	1	2,213	463	19.8	28.00
p150x2	2,048	512	1	1	2,563	799	19.8	28.37
p150x2	4,096	512	1	1	8,984	456	19.8	34.82
p150x2	8,192	512	1	1	19,680	416	19.7	45.59
p150x2	16,384	512	1	1	33,931	483	19.6	59.99
p150x2	32,768	512	1	1	67,812	483	19.4	94.18
p150x2	65,536	512	1	1	156,630	418	18.9	183.59
p150x2	130,048	512	1	1	380,812	342	18.1	408.97
p150x4	128	512	1	1	129	992	29.0	17.75
p150x4	1,024	512	1	1	1,232	831	28.8	18.98
p150x4	2,048	512	1	1	1,408	1,455	28.7	19.19
p150x4	4,096	512	1	1	4,975	823	28.6	22.85
p150x4	8,192	512	1	1	10,896	752	28.3	28.95
p150x4	16,384	512	1	1	25,643	639	27.8	44.05
p150x4	32,768	512	1	1	66,650	492	26.7	85.76
p150x4	65,536	512	1	1	194,450	337	24.9	214.98
p150x4	130,048	512	1	1	206,156	631	21.9	229.48

One run per row on each profile's mesh with vLLM 0.24.0, at one request at a time, 512 output tokens. On p150x2 every request had a unique cache salt; on p150x4 prefix caching is off, so every row is a cold prefill. Prefill tok/s/u is ISL divided by TTFT. Concurrency above 1 was not measured on either profile.

Limitations

• A single P150 is not supported: the model does not fit in its memory.
• Accuracy was measured on subsets (40 GPQA Diamond and 100 IFEval prompts) with thinking off, on p150x2 only. No agentic-coding benchmark was run. Speed was measured at one request at a time only, though p150x4 admits 8.
• Not for interactive use with long prompts on p150x2: a 32,768-token prompt takes about 68 s to first token, and a 130,048-token prompt about 381 s.
• On p150x4 prefix caching is off, so every turn of a multi-turn chat re-prefills the whole prompt.
• On p150x4, a 65,536-token prompt takes longer to first token than on p150x2 (194 s against 157 s).

Feedback

Questions or problems with this package: open a discussion at https://huggingface.co/tt-hous/laguna-xs-2.1/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:

component	built from
tt-metal	2d8292201660bff1ad9caf6cac556d9937f436d4 (dirty tree — the image includes uncommitted changes)
vLLM	v0.24.0
vllm-tt-plugin	c127c17d80d66ee83d23064d3a62ac844a1170de
code/ digest	5fcf91fffd6325bd (sha256, first 16 hex digits)
built	2026-09-14T13:55:33+00:00 by tt-model 0.1.0