Configura →
Modelli aperti

Tool open source e modelli open-weight, abbinati al dispositivo.

selbsai viene configurata attorno a limiti di memoria, termica e latenza. Per questo il configuratore ragiona per pacchetti di lavoro e fit modello/runtime, non per un unico nome marketing. Lo stack combina software open source classico con famiglie di modelli aperti e adatta il deployment al dispositivo, alle lingue, al workload e al profilo di licenza.

Private Assistant

Personal chat + light knowledge

Curated local assistant stack for fast private drafting and Q&A

Fast local assistance for private chat, email drafting, small document Q&A and day-to-day offline copilots.

RAM minima
16 GB
Velocità target
15-40 t/s
Office Knowledge

Documents + retrieval + coding support

Curated local knowledge stack with retrieval, longer context and workflow tools

Balanced local reasoning for document-heavy work, private search, research, coding support and office assistants.

RAM minima
32 GB
Velocità target
12-30 t/s
Coding + Agents

High-memory local agent node

Curated high-memory stack for coding agents, orchestration and heavier secure workloads

High-memory local intelligence for coding agents, larger private knowledge bases, orchestration and heavier secure workloads.

RAM minima
64 GB
Velocità target
8-18 t/s
Famiglie monitorate

Famiglie rappresentative che osserviamo davvero.

Queste famiglie servono a calibrare gli attuali tier di selbsai. I modelli effettivamente spediti possono cambiare quando nuove release aperte dimostrano migliori risultati su benchmark indipendenti.

Obsidian Personal · private assistant

Llama 3.1 8B Instruct

Meta Llama · Meta

Our compact baseline for fast local assistants, offline drafting and lighter retrieval on personal desktop devices.

Obsidian Core · office knowledge

Ministral 3 14B Instruct

Mistral / Ministral · Mistral

A strong local step-up for multilingual work, long-context tasks and richer instruction-following on balanced desktop systems.

Roadmap · multimodal local agents

Gemma 4 E4B / 26B A4B

Google Gemma · Google DeepMind

A model family we track for responsive multimodal chat, coding assistants and agentic workflows because Gemma 4 combines long context, native function-calling support and multi-token prediction drafters.

Obsidian Pro · coding and agents

Qwen3 30B A3B

Qwen · Qwen

One of the main higher-capability open families we watch for larger reasoning, coding and multilingual deployments on high-memory local devices.

Artificial Analysis

Independent model pages with direct comparisons across intelligence, speed, price, context window and methodology notes.

Hugging Face Open LLM Leaderboard

A widely used open-model benchmark hub for comparing community and lab releases across standard eval suites.

Arena Leaderboard

Useful for broad human-preference comparisons and keeping an eye on how major open releases stack up in live arena-style evaluation.

Cosa viene davvero installato

Non promettiamo che ogni unità Standard, Professional o Elite venga sempre spedita con lo stesso modello nominato. Un negozio serio di hardware AI locale deve scegliere lo stack open-weight giusto per il lavoro e aggiornare questa raccomandazione man mano che l’ecosistema migliora.

  • Il tier determina la scala di modello fattibile, la base di memoria e l’inviluppo di latenza.
  • I preset dei workflow determinano il tooling attorno al modello base: codice, revisione documentale, retrieval, operations o audit.
  • La selezione finale dipende dal mix linguistico, dalla sensibilità dei dati, dai vincoli di licenza e dal fatto che la build privilegi velocità, profondità o multimodalità.
  • Le posizioni nei benchmark cambiano nel tempo, quindi questa pagina rimanda a fonti vive di terze parti invece di fissare affermazioni di marketing destinate a invecchiare.
Curation layer

Hugging Face scale is useful only after filtering.

The value for customers is not simply that open models exist. The value is that selbsai turns a fast-moving model ecosystem into a controlled local setup with documented choices, workload fit, and a clear update channel.

Source reputation

Publisher history, release notes, model-card quality, community usage, and maintenance signals are reviewed before a model is treated as a provisioning candidate.

License and usage fit

The configurator now captures whether the customer wants permissive-only, commercial-ready, or restricted-model avoidance before final model selection.

Safe format preference

Where supported, selbsai prefers formats and runtimes with clearer supply-chain posture, including Safetensors, GGUF, MLX packages, and established local runtimes.

Hardware match

The selected model/runtime stack is checked against RAM, VRAM, thermal budget, storage, context length, and the customer's target workloads.

Provenance card

What the customer should know about the installed stack.

  • Model family, exact source repository, publisher, model-card link, and release reference.
  • Runtime path, file format, quantization level, checksum or verification reference where available.
  • License posture, intended use, known limitations, language fit, and benchmark references.
  • Selected update policy: stable, balanced, or fast track.

OCR and document extraction

For invoice, receipt and document-heavy presets we pair the language model with open OCR and document-understanding tooling rather than relying on the base LLM alone.

Retrieval and reranking

Search-heavy presets use additional embedding and reranking components so large local indexes stay usable at real-world scale.

Workload preimpostati

Software coding

Local help for private repositories.

Explain code, draft tests, review snippets, write scripts, and search repository notes without sending proprietary source to cloud tools.

  • Repo-aware Q&A
  • Test and script drafts
  • Error explanation

Documents and writing

Draft, rewrite, summarize, extract.

Create letters, policies, proposals, memos, summaries, and structured extracts from files that should stay inside your office.

  • PDF & DOCX ingestion
  • Memo and report drafts
  • Tables and summaries

Email and personal assistant

Inbox work without inbox exposure.

Draft replies, sort messages, extract tasks, prepare agendas, and turn notes into follow-ups from approved local exports.

  • Reply drafts
  • Action extraction
  • Meeting follow-ups

Research desk

Turns reading piles into briefings.

Compare sources, summarize PDFs, answer questions with citations, and prepare decision notes from local research folders.

  • Citation-aware Q&A
  • Long-context search
  • Briefing notes

Document review

Find clauses, risks, gaps, and dates.

Review contracts, policies, case files, leases, and due-diligence packs for obligations, inconsistencies, and missing attachments.

  • Clause search
  • Obligation extraction
  • Risk and gap lists

Sales assistant

Prepare better conversations faster.

Draft outreach, summarize accounts, prepare call notes, handle objections, and build proposals from approved sales material.

  • Proposal drafts
  • Call preparation
  • CRM-style summaries

Compliance management

Policies and evidence, searchable locally.

Answer audit questions, compare obligations, identify missing evidence, and prepare control summaries from internal policy folders.

  • Policy Q&A
  • Evidence checklists
  • Audit response drafts

Warehouse management

Operations support from local records.

Search SOPs, summarize shift notes, prepare supplier messages, and answer operational questions from warehouse documentation.

  • SOP search
  • Shift note summaries
  • Supplier message drafts

Inventory management

Stock lists, reorder issues, and reports.

Review stock exports, flag reorder risks, summarize item movements, and prepare plain-language inventory reports.

  • CSV and table review
  • Reorder flags
  • Inventory summaries

Company knowledge base

Ask your manuals, folders, and notes.

Build a local question-answer layer over manuals, procedures, project folders, email exports, and internal documentation.

  • Local vector index
  • Folder Q&A
  • Source-grounded answers