Ollama keeps backups of upgraded models so you can downgrade safely

3 days ago

Ollama v0.40.1 and v0.40.2 (October 7 and 8, 2026): models upgraded from earlier versions now keep a backup copy so downgrading stays safe, ollama list stops showing duplicates, ollama launch claude passes the full context window to Claude Code, and the local server forwards cloud usage and balance. Plus two Windows fixes. https://github.com/ollama/ollama/releases/tag/v0.40.2 https://github.com/ollama/ollama/releases/tag/v0.40.1 https://github.com/ollama/ollama/compare/v0.40.0...v0.40.2

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Upgraded models keep a backup copy on disk
  2. 0:14ollama list shows each upgraded model once
  3. 0:28ollama launch claude gives Claude Code the full context window
  4. 0:46The local server answers cloud usage and balance
  5. 0:59Two Windows failures stop, and the welcome screen skips a step
  6. 1:15Deleting the backups now means re-pulling to go back
Show transcript

Upgraded models keep a backup copy on disk

Ollama
Ollama
October 2026
Ollama
Upgraded models keep a backup copy on disk
An older Ollama can still run them. A future release deletes the backups for you.
Ollama release notes, Oct 8 2026 · Model upgrades

When Ollama upgrades a model you downloaded with an earlier version, it now keeps the original copy as a backup, so downgrading stays safe. Those backups take disk space until a future release removes them automatically.

ollama list shows each upgraded model once

Ollama
Ollama
October 2026
NAME                                ID              SIZE
gpt-oss:latest                      21fd97703fbe    13 GB
gpt-oss:latest                      4f44980319c2    13 GB
llamacpp:4f44980319c26335fc1cc5…    4f44980319c2    13 GB
PR #18874 · issue #18870

Those backups had been showing up in the model list. A model like gpt-oss showed up twice, plus a third entry named after a llama.cpp digest. The list and the tags API now hide the extra entries.

ollama launch claude gives Claude Code the full context window

Ollama
Ollama
October 2026

ollama launch claude passes the full context window

Claude Code’s own default
180K tokens
kimi-k2.6:cloud, as Ollama now passes it
262K tokens
PR #18855

For cloud models with a known context length, ollama launch Claude now tells Claude Code the model's real context window. Before, Claude Code assumed 180,000 tokens and compacted long sessions too early. A compaction window you set yourself is kept.

The local server answers cloud usage and balance

Ollama
Ollama
October 2026

Your local server reports cloud usage and balance

GET /api/usage
requests, tokens and cost over time
GET /api/balance
plan allowance left, plus purchased credit
PR #18829 · docs: /api/usage, /api/balance

If you use Ollama's cloud models, your local server now forwards two cloud endpoints for the signed-in account. Usage returns requests, tokens and cost over time, and balance shows what's left of your plan's allowance.

Two Windows failures stop, and the welcome screen skips a step

Ollama
Ollama
October 2026

Two Windows failures stop

  • Upgraded models open again instead of failing with “untrusted mount point”
  • Clef Flash answers instead of failing with “non-finite logit”
  • CLI welcome screen goes straight to the launcher, on every system
PR #18852 · PR #18777 · PR #18826

On Windows, upgraded models open again instead of failing with an untrusted mount point error, and Clef Flash stops failing every request with a non-finite logit error. Everywhere, pressing Enter on the welcome screen now skips the account step.

Deleting the backups now means re-pulling to go back

Ollama
Ollama
October 2026

Delete the backups now, and going back below 0.40 means a re-pull

“Model upgrades” in Ollama’s release notes on GitHub has the removal command. It needs jq. https://github.com/ollama/ollama/releases/tag/v0.40.2
Ollama release notes, Oct 8 2026 · Model upgrades

To get the space back now, the Model upgrades section of Ollama's release notes on GitHub has a command that deletes only the backups. It needs jq. After that, going back to a version older than 0.40 means pulling those models again.