Osmium

devstral:latest

Devstral on your own machine

Agent-oriented coder. Reach for it when Qwen 14B is too small and 32B will not fit.

Pull it

ollama pull devstral

Then pick it in Osmium under Settings → Local models. Nothing else to configure.

Parameters
24B
VRAM (4-bit)
16 GB
Disk
~12 GB
Context
128K tokens
Tool calling
Yes
Reads images
No
Generates images
No
Coding
8/10
Explaining
7/10

Scores are our own product guidance, not benchmark results.

Which hardware runs it

Checked against GPU memory. Osmium runs the same check on your real machine and also looks at system RAM and free disk. Unified-memory Macs need more headroom because the GPU shares one pool with the operating system.

  • 4 GB — GTX 1650 / MX570Needs 16 GB; this tier has 4 GB.Too big
  • 6 GB — RTX 2060 / 3050Needs 16 GB; this tier has 6 GB.Too big
  • 8 GB — RTX 3060 Ti / 4060Needs 16 GB; this tier has 8 GB.Too big
  • 12 GB — RTX 3060 / 4070Needs 16 GB; this tier has 12 GB.Too big
  • 16 GB — RTX 4060 Ti / 4080Runs, but memory is tight (16 GB against 16 GB).Tight
  • 24 GB — RTX 3090 / 4090Runs comfortably on 24 GB.Fits
  • Mac 16 GB unifiedRuns, but memory is tight (16 GB against 16 GB).Tight
  • Mac 24 GB unifiedRuns comfortably on 24 GB.Fits
  • Mac 36 GB unifiedRuns comfortably on 36 GB.Fits
  • Mac 64 GB unifiedRuns comfortably on 64 GB.Fits

Other coding models