for researchers

Local models. Your papers stay yours.

vinOS runs open-source LLMs on your GPU via Ollama. No paper uploaded to a server. No prompt logged by a third party. Reproducible: pin the model, pin the routine, share the .vinos/routines.yaml alongside your paper.

Download vinOS 2.0.5 See the recap routine →

Three outcomes for the research workflow.

Sovereignty, reproducibility, ambient synthesis. Your reading pile turns into notes overnight — on your hardware, with a model you control.

Sovereignty — no paper text leaves your machine.

Every routine that touches ~/Reading or your bib runs against a local Ollama model. Escalation to a frontier API is disabled by default for research routines.

route = "ollama" · enforced

Reproducibility — pin model + routine in git.

Model tag and routine TOML pinned by SHA. vinos-routine run prints the exact model version, quantization, and seed on every invocation.

model = "qwen2.5:32b-instruct-q4_K_M" · pinned

Nightly recap reads `~/Reading` for you.

Long-context local model summarizes new PDFs, cross-links to your bib, and emits spaced-repetition cards. Wakes you up with the synthesis, not the pile.

research-recap · 22:00 nightly

The proof — a routine + a model pull.

Two files. The routine (versioned in your paper's repo) and the exact ollama pull to reproduce the model behind it.

.vinos/routines/research-recap.toml toml

[routine]
name        = "research-recap"
schedule    = "0 22 * * *"                # nightly · 22:00 local
agent       = "recap"
route       = "ollama"                    # local only · never escalates

[model]
local       = "qwen2.5:32b-instruct-q4_K_M"   # 32B · 128k ctx · pinned
seed        = 42                              # for reproducibility

[tools]
allow       = ["read:~/Reading", "read:~/Zotero", "shell:pandoc"]

[prompt]
system      = """You synthesize the night's reading.
For each new PDF in ~/Reading since last run:
  1. Two-paragraph summary.
  2. Cross-links to prior papers in ~/Zotero (by DOI).
  3. Three spaced-repetition Q/A pairs.
Cite the model version and today's date at the bottom."""

[out]
render      = "markdown"
sink        = "~/Notes/recap-{{date}}.md"
open_on_login = true

[budget]
max_tokens  = 32000
max_dollars = 0.00                        # local · no dollar cost

~ · foot bash

# One-time: pull the 32B long-context local model.
$ ollama pull qwen2.5:32b-instruct-q4_K_M
pulling manifest ... 100%
digest: sha256:6c5f... · size: 19 GB

# Enable the routine.
$ vinos-routine enable research-recap
enabled · timer armed · next run: 22:00 today

# Verify what will run.
$ vinos-routine explain research-recap
routine: research-recap
model:   qwen2.5:32b-instruct-q4_K_M (local · sha256:6c5f...)
route:   ollama (frontier escalation disabled)
tools:   read:~/Reading, read:~/Zotero, shell:pandoc
budget:  32000 tokens · $0.00 · local-only

GPU + RAM sizing for local models: /models/. qwen2.5:32b at Q4 needs ~24 GB VRAM (or CPU + 32 GB system RAM at ~2 tok/s).

Questions researchers ask.

Can I cite the model version in my paper?

Yes — the model tag is pinned in the routine TOML (qwen2.5:32b-instruct-q4_K_M) and the SHA-256 of the manifest is printed on every run. Cite the tag, the SHA, and the routine's git commit; anyone with the same three can reproduce your synthesis bit-for-bit given the same input files.

Does this work with LaTeX / pandoc / bib tooling?

Yes. TeX Live, pandoc, biber, and Zotero (via zotero-cli) are in the standard bundle. Routines can invoke any of them as a shell: tool — the recap routine above uses pandoc to normalize PDFs before feeding them to the model.

GPU requirements for the big models?

Rough guide (full table on /models/): 7-8B quantized runs on 8 GB VRAM; 13B on 12 GB; 32B at Q4 needs ~24 GB (RTX 3090 / 4090 / A5000); 70B needs 48 GB+. CPU-only works for all sizes at reduced tok/s — fine for overnight routines, painful for interactive.

Can I share routines with collaborators?

Yes — drop .vinos/routines.yaml in your paper's repo and commit. Any collaborator who's booted vinOS runs vinos-routine load . to install the same set. Pair that with an ollama-models.lock file (list of tags + SHAs) for full reproducibility across machines.

How is this different from Elicit / Consensus / ResearchGPT?

Those are SaaS — you upload papers to their servers, they run their model, they log your prompts. vinOS is on your machine: your GPU, your model, your keys, your data. Cheaper long-term (no per-query fee), reproducible (pinned model version), and stays working if the vendor pivots or shuts down.

IRB / privacy-review implications?

Local-first materially simplifies most reviews — no third-party processor to disclose, no data-transfer agreement required, no subject-data leaving your machine. Print the routine TOML and the vinos-routine explain output as evidence of the pipeline.

Overnight synthesis, on your hardware.

Flash a USB, pull a model, enable the recap. Tomorrow morning the pile is a page of notes.

Download vinOS 2.0.5 Local model picker →

Uninstall: pacman -Rns vinos-*. No lock-in. MIT-licensed.