LM Studio news
Every thread that mentions this topic, across all sections
StrukLabs releases Harness Studio, a free open-source desktop app for local AI models, as an alternative to LM Studio and Ollama
A developer published the first public release of Harness Studio, an open-source desktop application positioning itself as an alternative to LM Studio and Ollama. The Rust/Tauri/Svelte app combines model management, an AI playground, inference server management, an API gateway, developer integrations, and performance monitoring in one tool. Released free under the MIT license via StrukLabs, the author is soliciting contributors and feedback to make it community-driven.
A developer published the first public release of Harness Studio, an open-source desktop application positioning itself as an alternative to LM Studio and Ollama. The Rust/Tauri/Svelte app combines model management, an AI playground, inference server management, an API gateway, developer integrations, and performance monitoring in one tool. Released free under the MIT license via StrukLabs, the author is soliciting contributors and feedback to make it community-driven.
Redditor proposes AI agents that write and inject mods into running games like GTA 5 and Skyrim in real time
Reddit user proposes AI agent that writes and injects game mods in real time
Reddit user proposes AI agent that writes and injects game mods in real time
GamePause: Windows tray app that unloads local LLMs from VRAM when gaming starts
GamePause is an open-source Windows tray utility for people who run local LLMs and game on the same PC. It detects game launches, unloads LM Studio/Ollama models to free VRAM, and restores them afterward. This removes the manual unload/reload chore when sharing GPU memory between models and games.
GamePause is an open-source Windows tray utility for people who run local LLMs and game on the same PC. It detects game launches, unloads LM Studio/Ollama models to free VRAM, and restores them afterward. This removes the manual unload/reload chore when sharing GPU memory between models and games.
Nyx AI: Android app with on-phone Qwen3.5 4B chat and document creation tools
Nyx AI is an Android app combining AI chat with tools to create Word, Excel, and PowerPoint documents plus interactive in-app previews. It runs an optional Qwen3.5 4B model directly on the phone, so users avoid cloud subscriptions and per-message provider fees. It suits Android users wanting local or self-chosen LLM access with office document generation.
Nyx AI is an Android app combining AI chat with tools to create Word, Excel, and PowerPoint documents plus interactive in-app previews. It runs an optional Qwen3.5 4B model directly on the phone, so users avoid cloud subscriptions and per-message provider fees. It suits Android users wanting local or self-chosen LLM access with office document generation.
Reddit user hits 30 tok/s running Tiel-Coder-35B-A3B Q4_K_XL on RX 6700 XT via llama.cpp Vulkan, seeks 40 tok/s
A Reddit user reports a month of llama.cpp Vulkan tuning to reach 30 tokens/second running Tiel-Coder-35B-A3B (Q4_K_XL and Q8) on an AMD RX 6700 XT with 12 GB VRAM, and asks the community how to reach 40 tps without changing quantization. The post includes a full LM Studio configuration, useful to practitioners tuning similar AMD/Windows setups.
A Reddit user reports a month of llama.cpp Vulkan tuning to reach 30 tokens/second running Tiel-Coder-35B-A3B (Q4_K_XL and Q8) on an AMD RX 6700 XT with 12 GB VRAM, and asks the community how to reach 40 tps without changing quantization. The post includes a full LM Studio configuration, useful to practitioners tuning similar AMD/Windows setups.
Developer releases Alpine Code, an MIT-licensed desktop coding agent with Python-function tools and hackable agent loop
A developer released Alpine Code, an open-source desktop coding agent for local models, along with the alpineagents Python library it is built on. The agent exposes a hackable harness — the agent loop, tools, and permission checks are all user-controllable — and runs against Ollama, LM Studio, vLLM, or any OpenAI-compatible endpoint. It matters to practitioners who want full visibility into agent internals without a terminal-based workflow or a hosted backend.
A developer released Alpine Code, an open-source desktop coding agent for local models, along with the alpineagents Python library it is built on. The agent exposes a hackable harness — the agent loop, tools, and permission checks are all user-controllable — and runs against Ollama, LM Studio, vLLM, or any OpenAI-compatible endpoint. It matters to practitioners who want full visibility into agent internals without a terminal-based workflow or a hosted backend.
AgenticOS: open-source, self-hosted browser-based platform for building AI agents, supports fully local model stacks
AgenticOS is an open-source (Apache-2.0), self-hosted platform for building AI agents in a browser, developed by Vstorm. It supports fully local setups using Ollama, vLLM, LM Studio, LiteLLM proxies, local embeddings, and local OCR, so no data leaves the user's network. It targets developers who want agent infrastructure without any third party in the chain.
AgenticOS is an open-source (Apache-2.0), self-hosted platform for building AI agents in a browser, developed by Vstorm. It supports fully local setups using Ollama, vLLM, LM Studio, LiteLLM proxies, local embeddings, and local OCR, so no data leaves the user's network. It targets developers who want agent infrastructure without any third party in the chain.
Cline Desktop v0.0.44 ships inline Mermaid diagrams, auto-reconnect, and fixes for local and Windows MCP providers
Cline shipped Desktop v0.0.44, adding interactive inline Mermaid diagram rendering in chat and automatic Cline Hub session reconnection. The release also fixes broken local/self-hosted providers, Windows MCP server loading, and custom Anthropic base URLs.
Cline shipped Desktop v0.0.44, adding interactive inline Mermaid diagram rendering in chat and automatic Cline Hub session reconnection. The release also fixes broken local/self-hosted providers, Windows MCP server loading, and custom Anthropic base URLs.
Reddit user running LM Studio/Ollama across AMD, NVIDIA, and Apple Silicon nodes seeks unified serving and monitoring setup
A Reddit user running local inference across three heterogeneous machines (AMD, NVIDIA, and Apple Silicon) asks what tooling others use for multi-node serving, routing, and observability. The post highlights a gap in existing self-hosted stacks like LM Studio and Ollama, which lack unified cross-node visibility into loaded models, token throughput, and memory pressure.
A Reddit user running local inference across three heterogeneous machines (AMD, NVIDIA, and Apple Silicon) asks what tooling others use for multi-node serving, routing, and observability. The post highlights a gap in existing self-hosted stacks like LM Studio and Ollama, which lack unified cross-node visibility into loaded models, token throughput, and memory pressure.
Building a local LLM/agent stack on a 128GB M5 Max Mac Studio for sensitive client data
A Reddit user buying a 128GB M5 Max Mac Studio asks which local LLM and agent stack to run for a consulting business handling sensitive client data. They want multi-step agentic capabilities — coding, browser automation, and local data analysis — without monopolizing all 128GB of unified memory, and are confused about how inference layers like MTPLX, LM Studio, and Ollama fit together.
A Reddit user buying a 128GB M5 Max Mac Studio asks which local LLM and agent stack to run for a consulting business handling sensitive client data. They want multi-step agentic capabilities — coding, browser automation, and local data analysis — without monopolizing all 128GB of unified memory, and are confused about how inference layers like MTPLX, LM Studio, and Ollama fit together.
Zero-dependency Go framework for running agent loops with local LLMs
A Go framework for building agent loops that run entirely against locally hosted models, with no external dependencies. It connects to local inference servers like Ollama, LM Studio, and vLLM. It targets Go developers who want to build LLM agents without cloud APIs or added dependency weight.
A Go framework for building agent loops that run entirely against locally hosted models, with no external dependencies. It connects to local inference servers like Ollama, LM Studio, and vLLM. It targets Go developers who want to build LLM agents without cloud APIs or added dependency weight.
Local LLM Cleaner — free Windows tool that finds duplicate AI model files and merges them via symlinks
A free Windows utility that scans for model files used by local AI apps like Ollama, ComfyUI, and LM Studio, then merges byte-identical duplicates into shared folders using symlinks so all apps keep working. It targets users running multiple local AI tools who lose tens of gigabytes to repeated model downloads. No LLM is used in the product itself; it manages models for running LLMs locally.
A free Windows utility that scans for model files used by local AI apps like Ollama, ComfyUI, and LM Studio, then merges byte-identical duplicates into shared folders using symlinks so all apps keep working. It targets users running multiple local AI tools who lose tens of gigabytes to repeated model downloads. No LLM is used in the product itself; it manages models for running LLMs locally.
Moka Chat gives Ollama models MCP tools via one npx command, with an inspector showing every tool call
A developer released Moka, a free, MIT-licensed local sandbox for testing prompts and MCP servers visually. It runs via 'npx @mokalabs/sandbox' with zero config, auto-detects running Ollama instances, and includes an inspector exposing raw JSON-RPC traffic, tokens, and latency. Developers can test tool calls and streaming, compare models side by side, and export working LangGraph Python scripts directly.
A developer released Moka, a free, MIT-licensed local sandbox for testing prompts and MCP servers visually. It runs via 'npx @mokalabs/sandbox' with zero config, auto-detects running Ollama instances, and includes an inspector exposing raw JSON-RPC traffic, tokens, and latency. Developers can test tool calls and streaming, compare models side by side, and export working LangGraph Python scripts directly.
How to Run an LLM on Your Own Computer With Ollama
A practical guide to running LLMs locally with Ollama
A practical guide to running LLMs locally with Ollama
SayWrite: Mac menu-bar dictation with local speech recognition and optional AI polish
SayWrite is a Mac menu-bar dictation app: hold Fn, talk, and text is typed at the cursor. Speech recognition runs locally; an optional AI Polish step uses Apple Intelligence or a local Ollama/LM Studio model to remove fillers and tidy rambling text. It targets Mac users who want private, per-app-styled dictation without uploading audio to the cloud.
SayWrite is a Mac menu-bar dictation app: hold Fn, talk, and text is typed at the cursor. Speech recognition runs locally; an optional AI Polish step uses Apple Intelligence or a local Ollama/LM Studio model to remove fillers and tidy rambling text. It targets Mac users who want private, per-app-styled dictation without uploading audio to the cloud.
llm-doctor: scanner for local LLM setups (duplicate weights, stale templates, context truncation)
llm-doctor is a diagnostic tool for people running local models. It scans model stores (Ollama, LM Studio, Hugging Face, MLX) for duplicate weights, stale chat templates, outdated GGUFs, and orphan blobs, and probes serving endpoints to catch silent problems like Ollama truncating context windows without erroring.
llm-doctor is a diagnostic tool for people running local models. It scans model stores (Ollama, LM Studio, Hugging Face, MLX) for duplicate weights, stale chat templates, outdated GGUFs, and orphan blobs, and probes serving endpoints to catch silent problems like Ollama truncating context windows without erroring.
WebBrain: open-source browser agent that runs locally with LM Studio/Ollama plus a 450M browser VLM
WebBrain is an open-source framework giving LLMs the ability to see and operate a browser. It avoids cloud lock-in by supporting local models via LM Studio and Ollama, as well as cloud APIs. The developer also trained webbrain-vl-2-450M, a small vision model fine-tuned on browser screenshots, so browser tasks need not hit large multimodal models.
WebBrain is an open-source framework giving LLMs the ability to see and operate a browser. It avoids cloud lock-in by supporting local models via LM Studio and Ollama, as well as cloud APIs. The developer also trained webbrain-vl-2-450M, a small vision model fine-tuned on browser screenshots, so browser tasks need not hit large multimodal models.
Is there any pruned version of Kimi K3 or any other top open MoE model whose download size is under 80GB, at or above Q4?
User seeks pruned MoE models under 80GB for low-VRAM laptop inference
User seeks pruned MoE models under 80GB for low-VRAM laptop inference
Llama.cpp and new model releases ...or why Great is the enemy of Good in the LLM world
A Reddit post argues llama.cpp's slow release cadence is ceding ground to faster forks. It cites GLM 5.3 Flash, which got zero-day support in Unsloth Desktop on release day (August 26) but still lacks llama.cpp main support a month later. The author warns users of prebuilt binaries like Ollama and LM Studio can't easily use forks, and suggests beta branches or standalone binary releases.
A Reddit post argues llama.cpp's slow release cadence is ceding ground to faster forks. It cites GLM 5.3 Flash, which got zero-day support in Unsloth Desktop on release day (August 26) but still lacks llama.cpp main support a month later. The author warns users of prebuilt binaries like Ollama and LM Studio can't easily use forks, and suggests beta branches or standalone binary releases.
PaperCat — Windows app that OCRs scanned mail and extracts deadlines, amounts and action items
PaperCat is a Windows desktop app for people who deal with lots of scanned paperwork. It OCRs documents locally, uses LLMs (OpenAI, Claude, Gemini, or local Ollama/LM Studio) to classify the document, extract deadlines and amounts, flag needed actions, and draft replies. It solves the problem of scanned mail ending up as an unorganized folder of PDFs.
PaperCat is a Windows desktop app for people who deal with lots of scanned paperwork. It OCRs documents locally, uses LLMs (OpenAI, Claude, Gemini, or local Ollama/LM Studio) to classify the document, extract deadlines and amounts, flag needed actions, and draft replies. It solves the problem of scanned mail ending up as an unorganized folder of PDFs.
Free GPU compatibility checker for Ollama, LM Studio, ComfyUI, PyTorch
A free web tool that checks whether a user's GPU (VRAM) and RAM are sufficient to run local AI tools like Ollama, LM Studio, ComfyUI, and PyTorch. It compares hardware specs against compatibility rules for each tool, answering a common question from people setting up local AI workloads. It is launched and the creator is seeking feedback.
A free web tool that checks whether a user's GPU (VRAM) and RAM are sufficient to run local AI tools like Ollama, LM Studio, ComfyUI, and PyTorch. It compares hardware specs against compatibility rules for each tool, answering a common question from people setting up local AI workloads. It is launched and the creator is seeking feedback.
Reddit users compare Unsloth Studio vs LM Studio after Unsloth Studio's day-one release
A Reddit user sparked discussion comparing Unsloth Studio, released that day, with LM Studio, arguing LM Studio has fallen behind amid rising new local-LLM platforms. The post questions LM Studio's focus on its BIONIC project and asks which tool practitioners prefer.
A Reddit user sparked discussion comparing Unsloth Studio, released that day, with LM Studio, arguing LM Studio has fallen behind amid rising new local-LLM platforms. The post questions LM Studio's focus on its BIONIC project and asks which tool practitioners prefer.
DeskPilot: open source Python desktop AI client for local LLMs with MCP and tool sandboxing
DeskPilot is an open source native desktop app for chatting with local LLMs through an agentic interface. It connects to Ollama, LM Studio, vLLM, or any OpenAI-compatible backend, and uses MCP with sandboxed tools to let models take actions locally. It targets users who want a full-featured, security-first desktop client without relying on cloud services.
DeskPilot is an open source native desktop app for chatting with local LLMs through an agentic interface. It connects to Ollama, LM Studio, vLLM, or any OpenAI-compatible backend, and uses MCP with sandboxed tools to let models take actions locally. It targets users who want a full-featured, security-first desktop client without relying on cloud services.
Preventing double-execution and tool replay in local LLM workflows (Rust tool gate)
A developer running local models via Ollama and LM Studio released v0.1 of an open-source Rust CLI/library, arcstone-mcp-sidecar, that prevents duplicate tool executions and replayed or mutated tool requests. The tool binds single-use claims to action, resource, and payload SHA-256, then atomically burns them on first execution. It targets a practical gap in local agent workflows where looping models retry or alter calls mid-sequence.
A developer running local models via Ollama and LM Studio released v0.1 of an open-source Rust CLI/library, arcstone-mcp-sidecar, that prevents duplicate tool executions and replayed or mutated tool requests. The tool binds single-use claims to action, resource, and payload SHA-256, then atomically burns them on first execution. It targets a practical gap in local agent workflows where looping models retry or alter calls mid-sequence.
Redditor builds two-agent verification layer for RAG/duckDB pipeline on self-hosted LM Studio server
A developer running a self-hosted AI server loaded with a few hundred files is planning two agents to audit their RAG pipeline: one to verify DuckDB's structured facts against source PDFs, and one to check that documents are faithfully retrievable. They ask whether the approach is worthwhile and whether existing agent repositories cover this, citing Hermes as their likely model choice.
A developer running a self-hosted AI server loaded with a few hundred files is planning two agents to audit their RAG pipeline: one to verify DuckDB's structured facts against source PDFs, and one to check that documents are faithfully retrievable. They ask whether the approach is worthwhile and whether existing agent repositories cover this, citing Hermes as their likely model choice.
PAM — open-source persistent memory system giving local AI assistants long-term recall
PAM (Persistent AI Memory) is an open-source memory system that lets AI assistants remember between conversations by storing memories, learning from them, and surfacing relevant context. It targets users of local/self-hosted LLM setups and works with OpenWebUI, LM Studio, Ollama, and MCP-compatible clients. Version 2.0 brings it to parity with the author's personal version.
PAM (Persistent AI Memory) is an open-source memory system that lets AI assistants remember between conversations by storing memories, learning from them, and surfacing relevant context. It targets users of local/self-hosted LLM setups and works with OpenWebUI, LM Studio, Ollama, and MCP-compatible clients. Version 2.0 brings it to parity with the author's personal version.
PromptDeck v1.1.0: open-source Tauri desktop app benchmarks local (Ollama, LM Studio) and cloud (OpenRouter, Groq, DeepSeek) LLMs side-by-side
The developer of PromptDeck released v1.1.0 of the open-source Tauri + React desktop app for prompting and benchmarking LLMs side-by-side, sharing it on r/LocalLLaMA. The new version adds cloud providers (OpenRouter, Groq, DeepSeek, Together, OpenAI, xAI and more) so local and cloud models can be compared in one run. It targets practitioners running local models via Ollama, LM Studio, and llama.cpp who want TPS/TTFT comparisons and scored test suites.
The developer of PromptDeck released v1.1.0 of the open-source Tauri + React desktop app for prompting and benchmarking LLMs side-by-side, sharing it on r/LocalLLaMA. The new version adds cloud providers (OpenRouter, Groq, DeepSeek, Together, OpenAI, xAI and more) so local and cloud models can be compared in one run. It targets practitioners running local models via Ollama, LM Studio, and llama.cpp who want TPS/TTFT comparisons and scored test suites.
Reddit user moving from RTX 3060 12GB desktop to RTX 5070 Ti laptop asks for local LLM setup tips
A Reddit user asks for configuration tips upgrading from a Ryzen 5 5600G desktop with RTX 3060 12GB and 16GB RAM to an Acer Predator laptop with Intel Ultra 9 275HX, RTX 5070 Ti 12GB, and 32GB RAM. They run local models via LM Studio in n8n document-generation workflows, and plan to keep using Claude Code for coding since the new hardware won't replace it.
A Reddit user asks for configuration tips upgrading from a Ryzen 5 5600G desktop with RTX 3060 12GB and 16GB RAM to an Acer Predator laptop with Intel Ultra 9 275HX, RTX 5070 Ti 12GB, and 32GB RAM. They run local models via LM Studio in n8n document-generation workflows, and plan to keep using Claude Code for coding since the new hardware won't replace it.
NYX AI — local-first Windows AI workspace for documents, files and code
NYX AI is a Windows desktop app that connects to local models via Ollama or LM Studio, or cloud providers, and applies AI tools to documents, spreadsheets, previews, files and code rather than just chat. It targets Windows users who want control over which model handles their data and where requests go. AppContainer isolation support adds a sandboxing option for AI file access.
NYX AI is a Windows desktop app that connects to local models via Ollama or LM Studio, or cloud providers, and applies AI tools to documents, spreadsheets, previews, files and code rather than just chat. It targets Windows users who want control over which model handles their data and where requests go. AppContainer isolation support adds a sandboxing option for AI file access.
Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro
Inco released Splash, an open-source inference engine built for Apple silicon, running Qwen3.8-27B at up to 144 tok/s on an M5 Max MacBook Pro. It claims up to 3× Ollama's decode speed and 2× oMLX, with nearly 4× gains when agents fan out into sub-agents. It ships as a CLI server plus an LM Studio runtime, targeting local agent work.
Inco released Splash, an open-source inference engine built for Apple silicon, running Qwen3.8-27B at up to 144 tok/s on an M5 Max MacBook Pro. It claims up to 3× Ollama's decode speed and 2× oMLX, with nearly 4× gains when agents fan out into sub-agents. It ships as a CLI server plus an LM Studio runtime, targeting local agent work.