Claude Desktop spawns 1.8 GB Hyper-V VM on every launch, even for chat-only use
TL;DR Highlight
Claude Desktop launches a ~1.8GB VM on every start, even for chat-only use.
Who Should Read
Developers using Claude Desktop only for chat who are bothered by its memory footprint on Windows or Mac
Core Mechanics
- Chat-only use still spawns a VM
- Every launch shows Vmmem at about 1,796–1,846 MB in Task Manager.
- That is over 11% of total memory on a 16 GB laptop.
- It starts after Cowork was used once
- Repro: Windows 11 with VirtualMachinePlatform enabled, use Cowork once, then restart.
- The vmcompute service also starts at boot via an RPC interface event.
- Killing processes doesn't stick
- After deleting 2,689 session files and killing vmcompute and vmwp, reopening the app respawned the VM immediately.
- Logs showed repeated 0xC037010D (Invalid JSON document) errors.
- Idle memory jumped before any use
- On a 16 GB system, idle usage rose from about 50% to about 62%.
- With normal app load, total reached 70–75%.
- Workarounds and their cost
- Disabling VirtualMachinePlatform stops the VM but also disables Cowork.
- Killing vmwp and vmcompute after each launch left chat working normally.
- What the reporter asked for
- Initialize the VM only when Cowork or agent mode is actually started.
- Clean up finished session data and keep chat working without VM infrastructure.
Community Reactions
- Same on macOS, per commenters
- Mac users reported a Virtual Machine Service for Claude using about 1.6–3 GB without touching the Cowork tab.
- One said Cmd+Q frees it, but relaunching reloads the VM within a minute.
- Config-file toggle reported
- A commenter suggested Developer > Open App Config File with "secureVmFeaturesEnabled": false.
- That comment received 4 confused (😕) reactions.
- Disabling the service worked for one user
- A commenter who doesn't use Cowork set CoworkVMService Start to 4 and vmmem stopped appearing while chat worked fine.
- The ~2 GB rootfs files were still downloaded.
- Pre-boot as intended design, some argued
- A commenter said early on the VM booted on Cowork queries but startup was slow, so it moved to pre-booting.
- An agent-harness developer said on-demand start with a 10-minute cooldown works better.
How to Apply
- Turn off the VM feature in app config
- If you don't use Cowork, open Developer > Open App Config File and set secureVmFeaturesEnabled to false under preferences.
- Restart the app and check Task Manager or Activity Monitor for the VM memory.
- Disable the service on non-Cowork PCs
- On Windows, try setting the CoworkVMService Start value to 4 if you don't use Cowork.
- Don't apply it if you plan to use Cowork.
- Start agent VMs only on demand
- If you build your own agent harness, start the VM on request and shut it down after 10 idle minutes.
- Keep working files on a mount so they survive VM shutdown.
Try It
# Paste into PowerShell (admin) — temporary workaround from the report: kill the VM processes (chat keeps working)
Stop-Process -Name vmwp -Force
Stop-Process -Name vmcompute -ForceTerminology
Related Papers
DeepInfer: High-performance inference framework for DeepSeek models
DeepSeek open-sourced DeepInfer, its production inference engine, under Apache-2.0.
Mercury 3: Diffusion at scale
Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality
DeepSeek banned from U.S. federal civilian agencies under new DHS order
DHS order bans DeepSeek from U.S. federal civilian agency IT systems
Timestral: Mistral's time-series foundation models
A 209M time-series model ranks first on three benchmarks and runs on CPU.
DeepSeek V4.1 Flash 1017 (Max)
DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.
If you don't run your own LLM proxy, how do you know whats going on?
Running your own LLM proxy shows which keys, models, and cache misses drive your usage.
DeepSeek: DeepSeek v4.1 Flash 1017
DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents
DeepSeek-v4.1-Flash-1017
DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length