Last updated: November 22, 2026
Ollama 0.40.17 is the current stable release, published November 10, 2026. It lets Ollama on Apple silicon serve several requests at once, adds Claude Code deferred tool loading to ollama launch claude, and adds the Inkling 2 Mini model. Version 0.40.16, released a day earlier, added Ripple Preview and fixed Responses API truncation.
Contents
- What is new in Ollama 0.40.16 and 0.40.17
- Parallel requests on Apple silicon
- How to update and turn on parallel requests
- ollama launch changes for Claude Code and other tools
- API and CLI fixes
- 0.40.17 compared with 0.40.15
- Limitations and tips
- Frequently asked questions
What is new in Ollama 0.40.16 and 0.40.17
Ollama shipped two releases in two days. The 0.40.16 release notes (November 9, 2026) and the 0.40.17 release notes (November 10, 2026) are the official record of what changed. No newer stable release was listed on GitHub on November 22, 2026.
This guide follows our earlier Ollama 0.40.15 guide to MiMo V2.6 and AutoTrust decision models and focuses on concurrency, ollama launch and API fixes.
| Release | Date | Headline changes |
|---|---|---|
| 0.40.16 | November 9, 2026 | Ripple Preview model; Qwen3.8 Flash Next uses about 27 GiB less memory on Apple silicon; /v1/responses truncation fix; CUDA 13.5 for Windows on ARM; updated llama.cpp |
| 0.40.17 | November 10, 2026 | Inkling 2 Mini (Apple silicon only); multiple requests at once on Apple silicon; deferred MCP tool loading for Claude Code; Vulkan and Windows integrated GPU fixes |
Ripple Preview, from Ricursive Intelligence, is a 65B dense model built for agentic coding (see our Ripple-65B Preview guide). Inkling 2 Mini, from Thinking Machines, is a 26B parameter model with 5.4B active that fits on a Mac at about 20GB; our Inkling 2 Mini local setup guide covers it in detail.
Parallel requests on Apple silicon
The 0.40.17 notes state: “Ollama on Apple silicon now serves multiple requests at once.”
The change comes from GitHub pull request #19197, “mlx: support concurrent generation”. It lets MLX models serve multiple generation requests at once through the OLLAMA_NUM_PARALLEL setting. Per the pull request, each request keeps its own cache and response stream, and long prompts are processed in 2048-token chunks interleaved with generation.
The pull request states that the default remained one request in that change. Ollama published these median timings on an Apple M5 Max with NVFP4 models, comparing serial handling with the new interleaved handling for two clients that submit together:
| Model | Second client starts streaming (before → after) | Both requests finished (before → after) |
|---|---|---|
| Gemma 4 12B | 7.45 s → 4.69 s | 9.57 s → 10.19 s |
| Qwen 3.8 27B | 13.84 s → 9.90 s | 17.23 s → 20.07 s |
| Nemotron 3.5 Lightning 30B-A3B | 3.31 s → 1.65 s | 4.99 s → 4.93 s |
The waiting client gets tokens sooner, but finishing both requests took longer on the two dense models. Single-client performance stayed close to the previous build.
A follow-up change, pull request #19785, was merged on November 20, 2026, after 0.40.17 shipped. It picks 1, 2, 4 or 8 slots from available memory when OLLAMA_NUM_PARALLEL is unset. Ollama’s Local FAQ already describes that behavior for MLX on Metal. On 0.40.17 itself, set the value explicitly if you want concurrency.
How to update and turn on parallel requests
- Update Ollama. On macOS and Windows, Ollama downloads updates automatically; click the menu bar or taskbar icon and select Restart to update. On Linux, re-run
curl -fsSL https://ollama.com/install.sh | sh. - Set the parallel limit. Start the server with
OLLAMA_NUM_PARALLEL=2 ollama serve. Raise it only if memory allows. - Size the context. The FAQ notes that parallel requests multiply the context allocation; a 2K context with 4 parallel requests results in an 8K context and more memory.
- Tune the queue if needed. When the server is overloaded it returns a 503 error.
OLLAMA_MAX_QUEUEdefaults to 512 queued requests, andOLLAMA_MAX_LOADED_MODELScontrols how many models stay loaded. - Watch memory. Run
ollama psto see which models are loaded and whether they sit on the GPU or CPU.
ollama launch changes for Claude Code and other tools
ollama launch sets up and starts coding tools against Ollama models. The Claude Code integration page shows the basic command, ollama launch claude, and a scripted form that uses --model and --yes. Changes in these releases:
- Deferred MCP tool loading (0.40.17):
ollama launch claudesupports Claude Code’s deferred tool loading, so MCP tools are loaded when needed instead of all at once. - Model routing (0.40.16):
ollama launch claudealways routes Claude Code to the selected model. - Retired cloud models (0.40.16):
ollama launchexplains when a cloud model has been retired. - OpenCode (0.40.17): thinking options are set from the model’s metadata.
API and CLI fixes
- Responses API truncation:
/v1/responsesrespectstruncationand no longer silently truncates by default (0.40.16). - Empty model lists:
/v1/modelsreturns an empty list instead ofnullwhen no models are installed (0.40.16). - Tool schemas: tool definitions passed through the API keep JSON schema keywords Ollama does not model, instead of dropping them (0.40.17).
- Thinking off: disabling thinking now maps to the model’s lowest supported reasoning level rather than being ignored (0.40.17).
- Scripts: running
ollamawith no arguments in a non-interactive shell prints help instead of an error (0.40.17).
On hardware, 0.40.17 makes Vulkan use mmap for integrated GPUs, fixing “Failed to allocate pinned memory” load failures, and Windows integrated GPUs no longer fail to allocate the KV cache for partially offloaded models.
0.40.17 compared with 0.40.15
| Area | 0.40.15 (November 3, 2026) | 0.40.16 and 0.40.17 |
|---|---|---|
| New models | MiMo V2.6; AutoTrust decision models | Ripple Preview; Inkling 2 Mini |
| Concurrency | Recurrent and vision GGUF models can serve more than one request | MLX models on Apple silicon serve multiple requests at once |
| ollama launch | Kimi Code CLI support; Pi model list filled in automatically | Deferred MCP tools for Claude Code; model routing fix; OpenCode thinking options |
| Other | Long-context (YaRN) scaling for Qwen3 and Qwen3.5 models | Responses API truncation fix; tool schema keywords preserved |
For background on earlier features such as speech and decision models, see our guide to Ollama’s October 2026 releases. For memory planning, see our explainers on the Ollama context window and Ollama system requirements.
Limitations and tips
- More memory per request: each active request keeps its own cache, and the pull request notes that idle slots retain memory and duplicate shared prompt prefixes.
- Lower throughput in some cases: speculative decoding is paused while multiple requests are active, which can reduce generation throughput.
- Inkling 2 Mini is Mac-only for now: Ollama says PC support will follow in an upcoming release.
- Other backends: the FAQ says non-MLX backends default to one parallel request unless you set
OLLAMA_NUM_PARALLEL.
Frequently asked questions
What is new in Ollama 0.40.17?
Ollama 0.40.17, released November 10, 2026, adds Inkling 2 Mini, lets Ollama on Apple silicon serve multiple requests at once, and adds Claude Code deferred tool loading to ollama launch claude. It also fixes integrated GPU loading on Vulkan and Windows.
How do I make Ollama handle parallel requests on a Mac?
Update to Ollama 0.40.17 or later, then start the server with OLLAMA_NUM_PARALLEL set, for example OLLAMA_NUM_PARALLEL=2 ollama serve. Each extra parallel request needs more memory, because every request keeps its own cache.
Does ollama launch claude load all MCP tools at startup?
Not since 0.40.17. Ollama’s release notes say ollama launch claude now supports Claude Code’s deferred tool loading, so MCP tools are loaded when needed instead of all at once.
How do I update Ollama to 0.40.17?
On macOS and Windows, Ollama downloads updates automatically; click the menu bar or taskbar icon and choose Restart to update. On Linux, re-run the install script: curl -fsSL https://ollama.com/install.sh | sh.
Did Ollama 0.40.16 change the Responses API?
Yes. According to the 0.40.16 release notes, /v1/responses now respects the truncation field and no longer silently truncates by default. /v1/models also returns an empty list instead of null when no models are installed.
Sources
- Ollama on GitHub: v0.40.17 release notes (November 10, 2026)
- Ollama on GitHub: v0.40.16 release notes (November 9, 2026)
- Ollama on GitHub: pull request #19197, mlx: support concurrent generation
- Ollama Docs: Local FAQ
- Ollama Docs: Claude Code integration
Image: Ollama logo by ParthSareen on behalf of Ollama, MIT License, via Wikimedia Commons. FreeAnswers is not affiliated with Ollama.