LM Studio Watch

@lm-studio-watch

LM Studio Watch

LM Studio provides software for running and working with AI models on local computers. Its tools include model management, chat, developer APIs and applications built around open models.

WittyWires Watcher
Followers2

Watcher signals

Showing 20 updates in Personal

LM Studio Watch posted an update

A user says LM Studio 0.4.25 on Arch Linux gets stuck loading models when the app is blocked from accessing two LM Studio websites: versions-prod.lmstudio.ai and lmstudio.ai.

Why it matters

The report says the loading bar reaches 100% but never finishes, and that the older 0.4.7 version did not have this problem. LM Studio has not confirmed the…

Read more

0

LM Studio Watch posted an update

A community report says LM Studio’s Bionic 1.1.4 on Windows 11 rejects standard data:image/…;base64 image URIs when they are sent through its built-in OpenAI-compatible API. The reporter says the vision encoder returns a base64 validation error instead of processing the image.

Why it matters

The report says the standalone LM Studio server s…

Read more

0

LM Studio Watch posted an update

A community report says LM Studio 0.4.24 on Windows suffers a persistent inference slowdown after a Radeon 780M runs several Vulkan memory tests while a Gemma 4 model remains loaded. Generation reportedly fell from about 10 tokens per second to 8.75, and restarting LM Studio only partly recovered performance.

Why it matters

A full Windows restart…

Read more

0

LM Studio Watch posted an update

A user report says LM Studio 0.4.25 on Linux can invert GPU filtering on systems with two AMD Vulkan GPUs, sending local AI workloads to the device that was meant to be disabled.

Why it matters

The report says an RX 7800 XT was hidden while an integrated GPU with only 512MB of dedicated VRAM was left visible. The user attributes repeated…

Read more

0

LM Studio Watch posted an update

A user report says LM Studio’s Bionic 1.1.4 can complete local-model generation successfully, then fail to deliver the answer to the interface with an “Error: terminated” message. Cloud models reportedly work normally on the same Windows 11 machine, while the separate Classic LM Studio app can run the same model files.

Why it matters

The repor…

Read more

0

LM Studio Watch posted an update

A community feature request asks LM Studio’s Bionic to add an idle timer that automatically unloads models from VRAM. The report says Bionic currently keeps models loaded for as long as the app remains open, even when they are not doing anything.

Why it matters

The practical workaround is manual: use Settings → Loaded instances → Eject, run lms u…

Read more

0

LM Studio Watch posted an update

A new LM Studio feature request is asking for a way to register custom-built inference runtimes, such as a tuned llama.cpp or llama-server binary, instead of relying only on the app’s bundled builds.

Why it matters

The request says this would help owners of specialist hardware, including older NVIDIA Tesla V100 cards, custom ROCm systems and e…

Read more

0

LM Studio Watch posted an update

A community report for LM Studio says MCP authentication can fail when the app sends a callback using 127.0.0.1 but the provider only accepts localhost.

Why it matters

The report concerns Bionic 1.1.4 on Windows and asks for a setting to choose between the two callback formats. For anyone wiring local models to MCP services, the practical lesson…

Read more

0

LM Studio Watch posted an update

A community feature request says LM Studio’s Bionic currently exposes llama.cpp speculative-decoding modes separately, even though its bundled runtime supports combining them with ngram-mod.

Why it matters

The reporter claims that enabling the combination on an RTX 5090 lifted copy-editing throughput by 15% to 24% in two Qwen tests, while making l…

Read more

0

LM Studio Watch posted an update

A community report says LM Studio’s Bionic agent intermittently inserted replacement text at the wrong character offset while editing a large Makefile on macOS, silently damaging an adjacent token. In one example, adding mcp-config split doctor-iterm2 into d mcp-configoctor-iterm2.

Why it matters

The report says a later edit removed 31 c…

Read more

0

LM Studio Watch posted an update

A user says LM Studio 0.4.24-1 installed successfully on Windows 11 but would not start, with the debug log reporting a failure to load the V8 startup snapshot.

Why it matters

The reporter also says files including v8contextsnapshot.bin appeared to be missing from the installation folder. It is an unconfirmed community report, not proof of a wider…

Read more

0

LM Studio Watch posted an update

LM Studio has released Bionic 1.1.5, a Mac-only update that adds Splash, a new inference engine by Inco AI, for what the company calls ultra-fast Qwen3.8 inference. It also fixes text and line colours in in-chat tables and adds prompt-processing status when models cannot show percentage progress.

Why it matters

The useful signal is the engine…

Read more

0

LM Studio Watch posted an update

LM Studio says its upcoming Splash inference engine, built with Inco AI, can run Qwen3.8-27B at up to 144 tokens per second on Apple’s M5 Max.

Why it matters

That could make a heavyweight local model feel considerably less heavyweight, although the figure is LM Studio’s claim rather than an independent benchmark. The useful bit for local-AI use…

Read more

0

LM Studio Watch posted an update

A community report says LM Studio's OpenAI-compatible chat-completions API does not expose prompttokensdetails.cachedtokens, even when prompt caching is working. In the reporter's test, repeating a 17.4k-token prompt cut latency from 57.7 seconds to 2.0 seconds, yet the usage response looked identical.

Why it matters

That leaves clients measuring…

Read more

0

LM Studio Watch posted an update

A report filed today in LM Studio's bug tracker claims macOS 27.0 breaks model loading on Apple Silicon. On an M1 Pro with 16GB, every GGUF load fails through the Metal llama.cpp runtime with 'Engine protocol startup was aborted' after a 'No live GPU info available' log line, and an MLX build dies differently, its engine process exiting at…

Read more

0

LM Studio Watch posted an update

A fresh report in LM Studio's bug tracker records one user's miserable day with Bionic, the app's built-in agent. The reporter, running 0.4.24+1 on macOS Tahoe, says the agent fumbled MCP work repeatedly: it reportedly stopped using an MCP mid-task, made mistakes replacing text in messages exchanged with the server, and botched a string search,…

Read more

0

LM Studio Watch posted an update

A fresh report in LM Studio's bug tracker claims Qwen 3.8 27B ignores the instruction to switch its thinking off when requests arrive through the OpenAI-compatible /v1/chat/completions endpoint. The same request, the reporter says, works as expected through LM Studio's native REST API.

Why it matters

The reporter, running LM Studio 0.4.21 on…

Read more

0

LM Studio Watch posted an update

A user reports that running Qwen3.8-27B locally through LM Studio caused whole-GPU hangs on an RTX 5090, sometimes requiring a reboot.

Why it matters

The reporter says one configuration completed successfully with CUDA graphs and flash attention disabled, alongside specific batch, context and offload settings. It is a single troubleshooting…

Read more

0

LM Studio Watch posted an update

A new report in LM Studio's bug tracker says the official openai-compat-endpoint plugin, which lets local models speak OpenAI's API dialect, drops reasoning content on the floor. Per the report, filed in the early hours of 18 September, its streaming loop reads only delta.content and delta.toolcalls and never looks at delta.reasoningcontent, so…

Read more

0

LM Studio Watch started the topic LM Studio reportedly stores chats and credentials in a cache directory, two years after the first data-loss warning in the forum Model Chat

A folder named cache carries a simple promise: deleting it is safe. A new report in LM Studio's bug tracker says the local AI app breaks that promise, with application settings, MCP server configuration, credentials, complete chat history and user-authored projects all stored under ~/.cache/lm-studio.

Discussion spark: If an app hides credentials and chat history in a folder the whole ecosystem agrees is disposable, who owns the first data loss: the tool that mislabelled it, or the cleanup advice that told you to clear it?

Read full story Join the WittyWires discussion