Releases·Tool Release
ollama releases v0.34.1 with MLX memory and cache fixes
Ollama published v0.34.1, a patch release updating its MLX runner and llama.cpp backends. Changes focus on memory handling: the mlxrunner now evicts prefix cache snapshots and waits for evicted runners and free system memory before loading the next model. The release also raises the token repeat limit to 100 and fixes a ChatGPT model selector spacing bug in the app.
14 engagement1 sourceMon, Sep 14, 2026, 10:14 PM
AI Overview
- ollama v0.34.1 makes mlxrunner check free memory and wait for evicted runners before loading the next MLX model
- v0.34.1 raises token repeat limit to 100 and returns an error instead of incomplete results
- v0.34.1 evicts prefix cache snapshots from the active conversation and scopes MLX array lifetimes instead of pinning and sweeping
- v0.34.1 keeps the gemma3n projector off the CPU and refreshes the app's ChatGPT model selector and Apps layout
Related
Releases·Tool Release
ollama/ollama v0.34.3: thinking controls in /api/show, MLX vision support
Ollama released v0.34.3, making GET /api/show advertise each model's thinking levels and default, using glm-5.3-flash:cloud as the example. The release also adds MLX-based Apple Silicon support for Nemotron H vision models and fixes macOS window reopening.
Ollama released v0.34.3, making GET /api/show advertise each model's thinking levels and default, using glm-5.3-flash:cloud as the example. The release also adds MLX-based Apple Silicon support for Nemotron H vision models and fixes macOS window reopening.
2 engagement1 sourcegithubSat, Sep 19, 2026, 12:02 AM
Products·launched
superfluid: local LLM daemon with Ollama/OpenAI/Anthropic APIs, persistent sessions and fair scheduling
superfluid is a local server daemon that fronts llama.cpp and MLX models behind Ollama-, OpenAI-, and Anthropic-compatible APIs so existing clients like Open WebUI keep working. It targets home-lab users running local models, addressing crash-lost conversations, chat latency when long agent prefills occupy the queue, and accidental exposure as an open inference endpoint.
superfluid is a local server daemon that fronts llama.cpp and MLX models behind Ollama-, OpenAI-, and Anthropic-compatible APIs so existing clients like Open WebUI keep working. It targets home-lab users running local models, addressing crash-lost conversations, chat latency when long agent prefills occupy the queue, and accidental exposure as an open inference endpoint.
2 engagement1 sourceredditThu, Oct 8, 2026, 03:30 PM
Releases·Tool Release
Local-llmup package lets users install or switch local LLM runtimes
A developer released Local-llmup, a package for managing local LLM installations and switching between runtimes such as Ollama, llama.cpp, MLX, and LM Studio. The tool aims to simplify runtime management for local AI workflows.
A developer released Local-llmup, a package for managing local LLM installations and switching between runtimes such as Ollama, llama.cpp, MLX, and LM Studio. The tool aims to simplify runtime management for local AI workflows.
74 engagement2 sourcesredditMon, Aug 10, 2026, 07:22 AM
Releases·Tool Release
Ollama ships v0.33.x releases with MLX runner updates and Qwen3.8 Flash Next support
Ollama released three versions in a week: v0.33.0 (Aug 21), v0.33.1 (Aug 26), and v0.33.2 (Aug 27). The updates add a Claude desktop app, MLX runner improvements including structured output support and Qwen3.8 Flash Next support, plus app and proxy fixes. Practitioners running local models on Mac benefit from Metal GPU timeout fixes and faster structured output.
Ollama released three versions in a week: v0.33.0 (Aug 21), v0.33.1 (Aug 26), and v0.33.2 (Aug 27). The updates add a Claude desktop app, MLX runner improvements including structured output support and Qwen3.8 Flash Next support, plus app and proxy fixes. Practitioners running local models on Mac benefit from Metal GPU timeout fixes and faster structured output.
Updated Wed, Sep 2, 2026, 12:11 AM — v0.33.2 released Aug 27 with dark mode restore, proxy catalog-change handling, and macOS handoff sync.
04036 engagement4 sourcesgithubFri, Aug 21, 2026, 10:52 PM
Social·Community
REAP/MTP Qwen3.8-Flash-Next MLX builds — here are the measured trade-offs to Jundot
A Reddit post by mensaprodigy announces three Qwen3.8-Flash-Next REAP builds for Apple Silicon, benchmarked on an M4 Max against Jundot's oQ4e MTP community package. All builds are qwen4_exp derivatives with 48 transformer layers, top-10 routing, a retained 512-expert native MTP predictor, and 262k configured context. The post gives coding scores, throughput, and memory footprints for each build, plus guidance on which to pick.
A Reddit post by mensaprodigy announces three Qwen3.8-Flash-Next REAP builds for Apple Silicon, benchmarked on an M4 Max against Jundot's oQ4e MTP community package. All builds are qwen4_exp derivatives with 48 transformer layers, top-10 routing, a retained 512-expert native MTP predictor, and 262k configured context. The post gives coding scores, throughput, and memory footprints for each build, plus guidance on which to pick.
24 engagement1 sourceredditMon, Sep 21, 2026, 06:54 PM