Been running Qwen2.5-32B-Instruct-Q5_K_M in ollama on a 3090 for weeks. Single Cursor window, ~18GB VRAM, snappy enough.
Opened a second project this afternoon so I could compare refactors. CubedLM died mid-edit with cudaMalloc failed: out of memory. Both windows were still "using" the same local endpoint.
Is anyone pinning one local model per workspace, or do you just pay Anthropic when you multitask?
Woke up to qwen2.5:14b instruct q4 K M answering like it had amnesia. Turns out yesterday's Ollama bump wiped the Modelfile override. Activity Monitor still showed 18GB resident so I kept yelling at…
spent friday night trying to rename a package with mlx community/Qwen2.5 32B Instruct 4bit on an M2 Max 64GB. got through the import graph, wrote a half finished pyproject.toml, then metal panic and…
Tried the new qwen2.5:14b instruct q4 K M in LM Studio on the Mac Studio last night. UI said "fits in 64GB" so I loaded it with a 32k context for a RAG eval. Ten minutes in, fans go nuclear, swap hit…
5 comments
Join the discussion
Log in to comment.
which qwen build exactly — the instruct Q5 or the coder one? i keep seeing people mix those and then blame the quant.
also did both windows share one ollama process or did you spin a second server? on my m2 i just crash the whole app if i try two contexts.
instruct Q5_K_M here too. both windows hit the same ollama — no second server.
what actually froze for me wasn't the model, it was the Cursor UI spinning while VRAM was already full. keyboard still worked in Figma which is how i noticed it wasn't a full machine lock.
Same pain. I got
CUDA error: out of memorythe moment Warp + Cursor both hit localhost:11434.I ended up putting ollama behind a tiny queue script so only one client talks at a time. Ugly, but cheaper than another $20/mo cloud seat for "just comparing branches".
the queue script approach is underrated. i tried
OLLAMA_NUM_PARALLEL=1and still got two contexts fighting for VRAM because Cursor keeps a ghost connection after you close the tab.killed the ghost with
ss -lptn 'sport = :11434'then it behaved. ugly demo tip but it saved a webinar once.I pin one model per project folder now.
.ollama/activesymlink + a Cursor rule that refuses to start ifnvidia-smishows another process on :11434.Paying Anthropic for the second window is still cheaper than another 3090. Hate that that's the math.