Ollama now runs Inkling 2 Mini, and Macs serve several chats at once
1 day ago
Ollama v0.40.16 and v0.40.17 add Inkling 2 Mini (Apple silicon only for now) and Ripple Preview, let Ollama on Apple silicon serve several requests at once, cut Qwen3.8 Flash Next's memory by about 27 GiB, load MCP tools on demand in ollama launch claude, and fix model loading on integrated GPUs.
Links
- Ollama v0.40.16 release notesgithub.com
- Ollama v0.40.17 release notesgithub.com
- Compare viewgithub.com
- Inkling 2 Miniollama.com
- Ripple Previewollama.com
- Concurrent MLX generationgithub.com
- MLX default parallelism from memorygithub.com
- Qwen3.8 Flash Next memorygithub.com
- Claude Code routing and retired modelsgithub.com
- Claude Code deferred tool loadinggithub.com
- Vulkan mmap on integrated GPUsgithub.com
- Windows iGPU pinned memorygithub.com
- Responses API truncationgithub.com
Ask about this presentation
Chapters
0:00Ollama now runs Inkling 2 Mini, a 26B model that fits on a Mac
0:14Ollama on Apple silicon now answers several requests at once
0:24Ripple Preview, a 65B model for agentic coding, runs with one command
0:33Qwen3.8 Flash Next now fits on a 128 GB Mac
0:47ollama launch claude loads MCP tools only when Claude Code needs them
0:58Integrated GPUs load models that used to fail, and long API inputs stay whole
1:12Upgrade now if you run Ollama on a Mac
Show transcriptHide transcript
Ollama now runs Inkling 2 Mini from Thinking Machines, a 26 billion parameter model that fits on a Mac in about 20 gigabytes. It reads images, calls tools, and solves about four in five real GitHub issues in a coding test. For now it needs Apple silicon.
On Apple silicon, Ollama now serves several requests at once. A second chat or agent used to wait for the first to finish. Left at the default, Ollama picks up to eight parallel slots from the memory you have free.
Ripple Preview, from Ricursive Intelligence, is a 65 billion parameter dense model built for agentic coding. Ollama run ripple preview pulls it and starts a session.
Qwen 3.8 Flash Next now uses about 27 gigabytes less memory on Apple silicon. It drops from roughly 98 to 71, which puts it within reach of a Mac with 128 gigabytes, at about the same speed.
Ollama launch Claude now supports Claude Code's deferred tool loading, so MCP tools load when they're needed instead of all at once. Claude Code always routes to the model you selected, and a retired cloud model gets an explanation.
On integrated GPUs, Vulkan now memory-maps the model, which fixes the failed to allocate pinned memory error. On Windows, partly offloaded models now get room for their cache. And the Responses API stops cutting long inputs short unless you ask it to.
Should you upgrade? I would say yes on Apple silicon, where Inkling 2 Mini and parallel requests run. On a PC with an integrated GPU, the loading fixes are reason enough, though Inkling waits for a later release. Ollama's release notes list every change.
Up next
1:21Ollama v0.40.15 runs Xiaomi's MiMo V2.6 and AutoTrust models that answer yes or no
@ollama
2 weeks ago
1:12Ollama v0.40.14 runs your models at their own temperature through the OpenAI API
@ollama
2 weeks ago
1:47Ollama now runs Liquid AI's LFM3 on Linux and Windows, speech included
@ollama
3 weeks ago
1:47Ollama now reads replies aloud with Phonon, and Flint reaches Linux and Windows
@ollama
4 weeks ago
1:13Ollama v0.40.8 runs Flint, Recursive's open coding model for agents, on Macs first
@ollama
1 month ago
1:45Ollama runs Perplexity's pplx-decider, a local 27B model that routes, checks and grades
@ollama
1 month ago
1:13Ollama v0.40.5 runs NVIDIA Nemotron 3.5 Bolt, a 74B vision model sized for one 24 GB GPU
@ollama
1 month ago
1:37Ollama run now answers yes-or-no questions with decision models
@ollama
1 month ago
1:31Ollama keeps backups of upgraded models so you can downgrade safely
@ollama
1 month ago
2:42Ollama v0.40.0: models run on MLX by default on Apple Silicon
@ollama
1 month ago
1:31Ollama v0.35.1: Clef decision models, now with images
@ollama
1 month ago
2:46Ollama v0.35.0: decision models that answer with choices and probabilities
@ollama
1 month ago