Model recommendations

What this is

A catalog of starting points for running LibraOS on a given piece of hardware, for a given task: which model to pull, which serving engine and recipe to run it with, which settings to set, what the licence requires you to check, and what has gone wrong before.

It carries no measured scores. A profile's status says whether we have validated it, not how well it performs. If you need measured, reproducible numbers for the LibraOS system layer itself, see Benchmarks; this catalog is a different thing — a map of "what runs, and how to start it," not "how good is the output."

Download the file directly:

starter-profiles.json

Fields

The file is a single JSON object.

Field	Meaning
schema_version	Integer schema version of this file.
as_of	Date this snapshot was produced.
libraos_release	The LibraOS release the snapshot was produced against.
notice	A caveat to show verbatim alongside any use of the data.
statuses	A map from each status value used below to a one-sentence explanation of what it means.
profiles	The list of starting-point profiles described below.

Each entry in profiles describes one model on one piece of hardware for one task:

Field	Meaning
profile_id	Stable identifier for this profile.
hardware	Machine-readable hardware identifier. Filter on this.
hardware_label	Human-readable hardware name.
deployment	Prose description of the deployment shape (for example, local model server vs. hosted gateway).
task	Machine-readable task identifier. Filter on this.
task_label	Human-readable task name.
model	The model weights to pull.
served_name	The model name/ID to configure the serving engine and LibraOS with, when it differs from model. May be null.
status	One of the keys in the top-level statuses map.
status_text	The explanation for status, copied from statuses for convenience.
serving	The serving engine, recipe, container image, and launch flags, or null where no serving configuration is published for this profile.
settings	A list of {key, value} environment settings to apply. May be empty.
licence	Licence notes for the weights and any flags that need attention (for example, a restriction on hosted use), or null.
red_flags	A list of {code, text} known problems with this profile. May be empty.
requirements	Hardware, system, runtime, external-service, and minimum-LibraOS-version requirements to check against the target machine.

Describe only what you read from the file: a future release may add fields, drop a profile, or change a status, and this page will not always be updated in lockstep. The notice field is the authoritative caveat; treat this table as a reading aid, not a substitute for it.

Starting points at a glance

Rows are hardware, columns are task. Each cell names the recommended_starting_point profile for that pair (served_name, falling back to model); the lines below it count the other profiles in that cell by status. A crashed profile is marked with a warning triangle. — means no profile is published for that pair yet. This table is generated from starter-profiles.json at build time, so it always matches the file above.

Hardware	Financial-filing Q&A, document attached
NVIDIA DGX Spark (GB10)	Qwen/Qwen3.8-Flash-Next 1× validated 5× not yet validated ⚠ 2× crashed on this hardware
Hosted model gateway	no recommended starting point yet 1× validation stale 1× not yet validated

All profiles

Hardware	Task	Model	Engine	Status
NVIDIA DGX Spark (GB10)	Financial-filing Q&A, document attached	Qwen/Qwen3.8-Flash-Next	vLLM 0.30 (myllmbox image v4, MTP K=5)	recommended starting point
NVIDIA DGX Spark (GB10)	Financial-filing Q&A, document attached	unsloth/Qwen3.6-35B-A3B-NVFP4	vLLM (mia-vllm-gb10-linear-b12x, MTP 2)	validated
NVIDIA DGX Spark (GB10)	Financial-filing Q&A, document attached	Mia-AiLab/Qwen3.8-Flash-Next-NVFP4	vLLM qwen38-flash-next image, MTP 3	not yet validated
NVIDIA DGX Spark (GB10)	Financial-filing Q&A, document attached	poolside/Laguna-S-2.1-NVFP4	vLLM 0.25.1 + DFlash	not yet validated
NVIDIA DGX Spark (GB10)	Financial-filing Q&A, document attached	Qwen/Qwen3.8-27B	SGLang (MTP)	not yet validated
NVIDIA DGX Spark (GB10)	Financial-filing Q&A, document attached	qwen/Qwen3.8-Flash-Next-TF	TensorFold v0.3.6.3 (MLX-format weights, MTP)	not yet validated
NVIDIA DGX Spark (GB10)	Financial-filing Q&A, document attached	nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4	SGLang + DSpark draft	not yet validated
NVIDIA DGX Spark (GB10)	Financial-filing Q&A, document attached	inclusionAI/Ling-3.0-flash-int4	—	crashed on this hardware
NVIDIA DGX Spark (GB10)	Financial-filing Q&A, document attached	RedHatAI/Muse-Glimmer-30B-NVFP4	—	crashed on this hardware
Hosted model gateway	Financial-filing Q&A, document attached	Qwen/Qwen3.6-35B-A3B	—	validation stale
Hosted model gateway	Financial-filing Q&A, document attached	zai-org/GLM-5.3-Flash	—	not yet validated

For AI agents

An agent can use this file directly, without logging in:

1. Download https://libraos.com/docs/catalog/starter-profiles.json.
2. Filter profiles by the target hardware and task identifiers (not the _label fields — those are for display).
3. Prefer the profile with status == "recommended_starting_point". Never pick one with status == "crashed_on_this_hardware".
4. Check that profile's requirements against the target machine (GPU vendor/family and memory, system architecture and OS, runtime engine, external services, and minimum LibraOS release) before recommending it.
5. Report as_of and libraos_release alongside the answer, and say plainly that the profiles carry no measured scores.

Example — list the profiles for one hardware/task pair:

curl -s https://libraos.com/docs/catalog/starter-profiles.json \
  | jq '.profiles[] | select(.hardware=="dgx-spark-gb10" and .task=="attachment_qa")
        | {profile_id, status, model, served_name}'

Output, from the file as of this writing:

{
  "profile_id": "dgx-spark-gb10.attachment_qa.gb10-qwen38-flash-next-hibrid48-vllm",
  "status": "recommended_starting_point",
  "model": "myllmbox/Qwen3.8-Flash-Next-hibrid48",
  "served_name": "Qwen/Qwen3.8-Flash-Next"
}
{
  "profile_id": "dgx-spark-gb10.attachment_qa.gb10-qwen36-35b-a3b-nvfp4-vllm",
  "status": "validated",
  "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
  "served_name": "unsloth/Qwen3.6-35B-A3B-NVFP4"
}
{
  "profile_id": "dgx-spark-gb10.attachment_qa.gb10-qwen38-flash-next-mia-nvfp4-vllm",
  "status": "not_yet_validated",
  "model": "Mia-AiLab/Qwen3.8-Flash-Next-NVFP4",
  "served_name": null
}
{
  "profile_id": "dgx-spark-gb10.attachment_qa.gb10-laguna-s21-nvfp4-vllm",
  "status": "not_yet_validated",
  "model": "poolside/Laguna-S-2.1-NVFP4",
  "served_name": "poolside/Laguna-S-2.1-NVFP4"
}
{
  "profile_id": "dgx-spark-gb10.attachment_qa.gb10-qwen38-27b-nvfp4-sglang",
  "status": "not_yet_validated",
  "model": "RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead",
  "served_name": "Qwen/Qwen3.8-27B"
}
{
  "profile_id": "dgx-spark-gb10.attachment_qa.gb10-qwen38-flash-next-mlx4-tensorfold",
  "status": "not_yet_validated",
  "model": "Vontra/Qwen3.8-Flash-Next-MLX-4bit-MTP",
  "served_name": "qwen/Qwen3.8-Flash-Next-TF"
}
{
  "profile_id": "dgx-spark-gb10.attachment_qa.gb10-nemotron35-lightning-nvfp4-sglang",
  "status": "not_yet_validated",
  "model": "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4",
  "served_name": "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4"
}
{
  "profile_id": "dgx-spark-gb10.attachment_qa.gb10-ling30-flash-int4-sglang",
  "status": "crashed_on_this_hardware",
  "model": "inclusionAI/Ling-3.0-flash-int4",
  "served_name": "inclusionAI/Ling-3.0-flash-int4"
}
{
  "profile_id": "dgx-spark-gb10.attachment_qa.gb10-muse-glimmer-30b-nvfp4-vllm",
  "status": "crashed_on_this_hardware",
  "model": "RedHatAI/Muse-Glimmer-30B-NVFP4",
  "served_name": null
}

A Python equivalent:

import json
import urllib.request

with urllib.request.urlopen("https://libraos.com/docs/catalog/starter-profiles.json") as r:
    data = json.load(r)

matches = [
    p for p in data["profiles"]
    if p["hardware"] == "dgx-spark-gb10" and p["task"] == "attachment_qa"
]
recommended = next((p for p in matches if p["status"] == "recommended_starting_point"), None)

Once an agent has resolved an upstream model this way, pointing LibraOS at it is the same step covered in Model settings — OPENAI_API_BASE, OPENAI_MODEL, and the related LIBRA_OS_* slots. A profile's settings list gives the exact keys and values for that profile.

Inside a LibraOS installation

The Desk Developer page shows these same starting points matched against the machine LibraOS is actually running on. If the operator has also configured a measured catalog, the Developer page additionally shows benchmark figures for that installation — this public file never does.

Agents running inside an installation can call the built-in models_recommend tool instead of fetching the public file, or call the API directly:

• GET /v1/managed/models/recommendations — list profiles, with a LibraOS user or service-client token.
• GET /v1/managed/models/recommendations.md — the same list as an agent-readable guide.
• GET /v1/managed/models/recommendations/{profile_id} — a single profile.

Updating

This file is refreshed with each LibraOS release. Check as_of and libraos_release before relying on it, and re-fetch rather than caching it across releases.