Skip to content

πŸ—οΈ Target architecture recommendation

The right architecture is not the purest on day one. It is the one that lets you validate the workflow without lying about sovereignty.

Step 1 β€” Practical MVP

To learn fast:

  • Cursor CLI or Aider;
  • manual run;
  • Markdown report;
  • dedicated Git branch;
  • human validation.

Cursor CLI (owned by SpaceX since August 20261) is very productive for testing the idea. Aider is closer to the sovereign target because it can call Ollama directly, but its development is frozen (see its fiche).

Step 2 β€” Controlled runner

To automate:

  • scheduled task (cron, systemd timer, self-hosted GitHub Actions, or Agent Canvas automation on an internal backend2);
  • dated branch;
  • run logs via vault-log-run under .agents/vault-maintenance/runs/;
  • source report;
  • notification without automatic merge.

Step 3 β€” Sovereign target

Recommended stack:

LayerRecommended choiceRole
Code agentOpenHands (CLI/SDK) β€” or Aider (frozen, see fiche) for experimentsModifies files and works with Git
Local modelOllama or vLLM + 27–30B agentic model (Qwen3.8-27B, Muse Glimmer 30B) or specialized coderOn-prem inference with sufficient reasoning
GatewayLiteLLM (lexicon) β€” monthly update mandatory (vulnerabilities exploited in 2026, one month of support per minor line)34OpenAI-compatible API, routing, logs
SearchSearXNGSelf-hosted web search
Runner and sandboxOpenHands / Agent Canvas25Docker agent, scheduled automations, orchestration of third-party agents via ACP
Fast MVPCursor CLIInitial productivity

Minimum sizing for local Aider

For a realistic sovereign target:

UseRecommended local modelHardware reading
Simple suggestions, small filesSpecialized coder 7B/8Buseful to learn, not reliable as an autonomous agent
Controlled fixes on Markdown vaultDense 14B coder (Qwen2.5-Coder-14B, from 2024; the Qwen3-Coder range has no 14B)6practical floor, with strict human validation
Regular maintenance, multi-file auditCoder 24–32B or higher: Devstral Small 2 (24B, December 2025), the Qwen3-Coder-30B-A3B MoE, or Qwen2.5-Coder-32B (2024)6recommended target if the agent must produce usable diffs
Agentic work with long context, 24 GB VRAMAgentic generalist 27–30B such as Qwen3.8-27B (~18 GB in Q4, 262k context) or Muse Glimmer 30B in Q4 2026; the Qwen3.6-35B-A3B MoE (3B active) remains the one recommended by the OpenHands docs (May 2026)context β‰₯ 32k tokens7; see Choose your model
Large refactor or long reasoning32B+ with large context, or frontier non-sovereign model in MVPsovereignty vs quality trade-off

Key point: the agent that acts on files needs more reasoning than the assistant that retrieves information. VRAM budget must be sized for the editing model, not only the chat model.

Concrete recommendation for this vault

  1. Short term: continue with Cursor/Aider under human validation.
  2. Medium term: OpenHands (CLI, or Agent Canvas with scheduled automations and a local LLM profile) or Aider as long as it works, + Ollama + recent agentic model (in Q4 2026, a 27–30B generalist such as Qwen3.8-27B or Muse Glimmer 30B; or the Qwen3.6-35B-A3B MoE recommended by OpenHands in Q2 2026) + SearXNG + maintenance scripts27.
  3. Long term: LiteLLM as gateway (maintained minor line, patched within the month)4, vLLM if throughput is needed, OpenHands/Agent Canvas for automations and complex sandboxed tasks25.

See also

πŸ“š Sources and References

Footnotes

  1. Cursor, Cursor is now a part of SpaceX, 2026-08-14. https://cursor.com/blog/joining-spacex ↩

  2. OpenHands, Introducing Agent Canvas (self-hosted control center, scheduled or event-driven automations, LLM profiles, local/Docker/VM/Kubernetes backends, MIT license), 2026-06-16. https://www.openhands.dev/blog/introducing-agent-canvas ↩ ↩2 ↩3 ↩4

  3. BerriAI, GHSA-7hp6-4w63-5g45 (internal_user β†’ proxy_admin β†’ host execution escalation, CVSS 9.9, fixed in 1.100.4 / 1.101.3 / 1.102.2 / 1.103.1), 2026-09-30; CISA, Known Exploited Vulnerabilities Catalog (CVE-2026-42208, CVE-2026-42271, CVE-2026-59822), catalog dated 2026-10-08. https://github.com/BerriAI/litellm/security/advisories/GHSA-7hp6-4w63-5g45 Β· https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json ↩

  4. LiteLLM, Version Support Policy (since 2026-06-29, only the four most recent stable minor lines receive fixes), 2026-06-20. https://docs.litellm.ai/blog/version-support ↩ ↩2

  5. OpenHands, Use any coding agent in OpenHands with ACP (Agent Client Protocol: Claude Code, Codex, Gemini CLI; ACPAgent in the SDK), 2026-06-18. https://www.openhands.dev/blog/use-any-coding-agent-in-openhands-with-acp ↩ ↩2

  6. Qwen, Qwen2.5-Coder-32B-Instruct (family from 0.5 to 32B, 14B and 32B under Apache 2.0, technical report arXiv:2409.12186 of September 2024); QwenLM, Qwen3-Coder (README: 30B-A3B, Next and 480B-A35B, no dense 14B or 32B variant), accessed 2026-11-19; Mistral AI, Introducing Devstral 2 and Mistral Vibe CLI (Devstral Small 2: 24B, Apache 2.0; Devstral 2: 123B, modified MIT), 2025-12-09. https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct Β· https://github.com/QwenLM/Qwen3-Coder Β· https://mistral.ai/news/devstral-2-vibe-cli ↩ ↩2

  7. OpenHands Docs, Local LLMs (Ollama, vLLM, SGLang, LM Studio; Qwen3.6-35B-A3B recommended, β‰₯ 24 GB of VRAM when quantized, 32k context), updated 2026-05-21. https://docs.openhands.dev/openhands/usage/llms/local-llms ↩ ↩2

Give my opinion

Everything is optional: one answer is enough. This site sends nothing: you send it, the way you choose.

Your noteNo note
Your opinion is about
About what? (several choices possible)
What will be sent
Send by e-mail

Private, no account: opens your mail program with the message ready, for ia-on-prem@damien.becherini.fr.

Open a GitHub issue β†—

Public, with a GitHub account. You read the issue again before you publish it.

To paste it wherever you want.

Give at least a note, a subject or a message to be able to send.