Opencode + lmstudio : first prompt very slow
The report surfaced without fanfare on April 13th: a user running Opencode alongside LM Studio for local inference noticed something specific enough to be worth documenting. Not general sluggishness, not the ambient frustration of underpowered hardware — a cold-start penalty, precise and reproducible. The first prompt of any session arrives slowly. Everything after it behaves normally. That pattern matters because it points away from model size or compute constraints and toward something in the initialization sequence — how the two tools hand off context at session start, how LM Studio loads or fails to pre-warm the model, or how Opencode structures its first-call payload. The post scored 4.2 on LocalLLaMA, modest engagement but from a community that reads stack traces for fun and rarely confuses latency with impatience. Since that single report, nothing has moved. No follow-up configurations, no corroborating accounts, no response from either the Opencode or LM Studio teams.
If confirmed, here is what this means. Opencode is positioning itself as a viable local-first coding assistant — the kind of tool that keeps sensitive codebases off cloud infrastructure. Cold-start latency is a friction point that sounds minor until you factor in developer workflows: the first prompt of a session is almost always the most important one, the framing question, the context-setter. A five-second wait is tolerable. A thirty-second wait breaks the thought. If this is a consistent, configuration-independent problem rather than a hardware-specific quirk, it undermines the local inference pitch at its most visible moment. For LM Studio, it raises a quieter question about how well it exposes pre-warming controls to the tools sitting above it in the stack.
Watch for other LocalLLaMA users reporting the same cold-start signature across different hardware configurations, or for either team acknowledging the issue in their public changelogs or Discord channels — either would shift this from anecdote to pattern.
<!-- SC_OFF --><div class="md"><p>I run pretty large models (DS4 Flash, Qwen3.8 Flash, ...) on a machine with pretty slow prompt processing (M2 Ultra).</p> <p>The first prompt
<!-- SC_OFF --><div class="md"><p>Model size gets a lot of attention. I believe the real difference between a frustrating and a great local LLM experience comes from everything
<!-- SC_OFF --><div class="md"><p>Curious to hear how long you guys are waiting for a long context session(100k tokens+) to resume in coding agents running locally. </p>
<!-- SC_OFF --><div class="md"><p>I have experience using llama.cpp on Windows/Linux with 8GB NVIDIA card (384 GB/s bandwidth) and offloading to CPU to run MoE models. I typically
<!-- SC_OFF --><div class="md"><p>Recently I started using LMStudio to load local models and use them with ClawdBot, when I started using it I could offload 100% of the model
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1w44dj9/how_to_debug_prompt_caching_issues_in_omlx/"> <img alt="How to debug prompt caching issues in omlx?"
<!-- SC_OFF --><div class="md"><p>I've been having decent success with some local models, but I've had a bit of an issue when it comes to capabilities with knowledge and/or the
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1sxwjfh/ive_got_a_feeling_that_llamacpp_is_not_the/"> <img alt="I've got a feeling that Llamacpp is not the
NewsHive monitors these sources continuously. All signal titles above link to the original reporting.
Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →