Ollama 0.40.6 loads Q2_0 GGUFs: fix for Bonsai 2 `size overflows`, plus Clef on cloud and a Windows runner cleanup

Ollama v0.40.6 (October 16, 2026) teaches its GGUF reader the Q2_0 tensor and file types, so stock Bonsai 2 Q2_0 packs stop failing with unsupported tensor size overflows. The same release adds Cloudflare's Clef decision model on Ollama Cloud and kills orphaned Windows runners via Job Objects.

If you tried to import Prism ML’s Bonsai 2 Q2_0 GGUFs into Ollama and hit unsupported tensor "…" size overflows, v0.40.6 is the fix. The release, published October 16, 2026, teaches Ollama’s GGUF reader the Q2_0 tensor and file types that upstream llama.cpp already ran, and it also ships two smaller builder-facing changes: Cloudflare’s Clef decision model on Ollama Cloud, and a Windows Job Object so orphaned runners die with the parent process.

Builder question this answers: why does ollama create / ollama run reject Bonsai 2 Q2_0 files that llama.cpp already loads, and what else landed in 0.40.6?

The Q2_0 gap, and what 0.40.6 closes

PR #18967 (merged October 15, 2026) states the mismatch clearly: llama.cpp b11669 (bundled in Ollama 0.40.5) can run Q2_0 GGUFs, but Ollama’s Go GGUF reader did not know tensor type 42 / file type 41. Create and load then failed with unsupported tensor "token_embd.weight" size overflows. The PR adds Q2_0 as 64 weights, 18 bytes per block, and the author reports Apple Silicon create/run success for Bonsai 2 2B / 4B / 9B / 27B with thinking, vision, and tools — the same files that failed on 0.40.5.

That matches Prism ML’s Formats & Runtime Support guide: Bonsai 2 ships three GGUF packings, and only upstream Q2_0 is meant for stock llama.cpp / apps that bundle it. PQ2_0 and PTQ1_0 stay PrismML-binary-only. Prism also warns that a runtime which knows the file type but not Bonsai 2’s Hadamard activation transform can load Q2_0 and still emit gibberish — so if text looks wrong after upgrade, check that your Ollama build includes a llama.cpp at or after b11555, not only the type table.

The library tag bonsai-2:27b-q2_0 is the practical entry point once you are on 0.40.6:

ollama run bonsai-2:27b-q2_0

Clef on Ollama Cloud

PR #19036 documents that Clef is now available on Ollama’s cloud. Ollama’s library page describes Clef as a 27B multimodal decision model from Cloudflare, fine-tuned from Qwen3.8-27B, that turns a state plus a schema of typed questions into decisions. Tags include clef:latest, clef:27b, clef:cloud, and clef:27b-cloud, with a cloud badge on the listing and a note that Clef needs Ollama 0.35.1 or later.

Windows: no more orphaned runners

PR #18500 assigns runner processes to a singleton Windows Job Object with JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE. When ollama serve exits — including forced kills where graceful cleanup cannot run — Windows terminates owned llama-server and MLX subprocesses. Desktop users were mostly fine already because the app shuts the server down cleanly; CLI users who relied on non-graceful taskkill / Stop-Process are the ones who saw zombies.

What to do

Upgrade to Ollama v0.40.6 (or newer).
For Bonsai 2 on stock Ollama, use the Q2_0 packs (or the library tags), not PQ2_0 / PTQ1_0.
If you still see nonsense output after a successful load, treat it as a Hadamard/runtime-version problem per Prism’s docs, not a corrupt download.
On Windows CLI hosts, re-test that killing ollama serve no longer leaves runners behind.

For the image-side Bonsai family, see our Bonsai Image 4B local guide and the earlier Bonsai Image 4B release note. For low-bit decode on NVIDIA via llama.cpp, see b11713’s Q2_K / Q3_K CUDA speedups.