Ollama “Downloaded in a Format That Is No Longer Supported”: A Confusing Error, Clearly Fixed

Short answer: since Ollama 0.40.4, the System One endpoint (/v1/systemone) refuses decision models stored in a format that is no longer supported. If the model is an official one such as nimble, tev1 or clef, run ollama pull again and check that its ID changes. If it is a third-party or self-built GGUF that declares CAPABILITY decision, re-pulling will not help: Ollama now needs a GGUF with its own decision head, and there is no workaround in current releases. We reproduced the message on Ollama 0.40.12 on October 29, 2026 (Beijing time).

The exact message

The API returns HTTP 400 with this body:

{"error":"'nimble:latest' was downloaded in a format that is no longer supported. To re-download, run 'ollama pull nimble:latest'"}

It appears on decision requests: calls to System One, the decision API in the official decision docs, and ollama run --ask. In our test, the CLI printed the same text after Error:, while a plain /api/chat request to the same model still answered.

Why Ollama says the format is no longer supported

In 0.40.4, Ollama changed how decision models run. An Ollama maintainer explained it in issue #18998: before 0.40.4, System One used a custom implementation with models configured for that engine. 0.40.4 switched to native llama.cpp support, the official models were updated for the new engine, and this message was added to tell users who still had the old copy.

The check itself is short. In server/routes.go (v0.40.12), the systemOneLlamaServer function reads the decision.type key from the GGUF:

• If decision.type exists, the model runs on llama.cpp’s decision heads.
• If it is missing and the model declares the decision capability, Ollama returns the format that is no longer supported message.
• If it is missing and there is no decision capability, the message is model "" is not a llama.cpp decision model.
• Old Clef GGUFs built on the qwen35 architecture also get the format message, because llama.cpp now expects the clef architecture.

Our test on Ollama 0.40.12

We wanted to see the message ourselves without waiting for an old official model. So we built one that fails the check on purpose: a small ordinary GGUF (SmolLM2-135M-Instruct, Q4_K_M) with a two-line Modelfile, FROM ./smol.gguf and CAPABILITY decision. The Modelfile docs warn against this: declare decision only for weights trained for System One’s decision prompts.

• ollama create succeeded.
• POST /v1/systemone returned HTTP 400: 'smoldecision:latest' was downloaded in a format that is no longer supported. To re-download, run 'ollama pull smoldecision:latest'.
• Following the advice, ollama pull smoldecision:latest failed with Error: pull model manifest: file does not exist. The model was never in a registry, so there is nothing to re-download.
• The same GGUF without CAPABILITY decision gave model "smoltest" is not a llama.cpp decision model. /api/chat requests to either model still returned normal text.

So the text “downloaded in a format” is a little misleading. What Ollama actually checks is the file’s metadata, not how it got onto your disk.

How to fix it

Official decision models: re-pull and compare the ID

This is the case in #18998. The user upgraded to 0.40.4 and nimble, tev1 and clef-flash stopped working. A maintainer asked for ollama list nimble:latest, pointed out that the local ID did not match the library ID, and asked for a fresh ollama pull nimble:latest. After that, all three models worked on 0.40.4.

1. Run ollama list and note the ID.

2. Run ollama pull .

3. Run ollama list again. The ID should be different. If it is the same, the pull did not replace the old copy. In #18998 the user’s first re-pull attempt was also reported as “did not work”; a later pull returned a new ID and fixed it.

4. Retry the System One call. It should no longer return format that is no longer supported.

Third-party or self-built GGUF decision models: no fix yet

Issue #19292, still open on October 29, 2026, reports this case. Its author built a decision model on an ordinary language-model head, published it as aminroudaki/decisio-gemma:q4_k_m, and says it works on 0.40.3 but returns this error on 0.40.4, 0.40.7 and 0.40.9. Re-pulling does not help because the published file is the one being refused. The issue asks Ollama to support such models again. Until that happens, your options are:

• Use an official decision model. The docs list Nimble, Tev1, Clef and Clef Flash and say System One needs Ollama v0.35.0 or newer.
• Stay on the version the issue says still works (0.40.3) for that one model. The Linux docs show how to install a specific version. You give up every fix since then, so treat it as temporary.
• If plain text answers are enough, call the model through /api/chat. That worked for our test model; whether it suits a model trained only for decisions is your call.

Removing CAPABILITY decision does not get around the format that is no longer supported check. As our test showed, you just get the other refusal.

Quick answers

Does format that is no longer supported mean my model file is damaged?

No. The file is readable. It lacks the decision metadata that Ollama 0.40.4 and later require.

Do I need to re-pull models after every Ollama update?

Not in general. In #18998 the maintainer said that once the models were updated for 0.40.4, there was nothing else to do.

Where can I read about these models?

Our Liquid AI d1 overview covers Liquid AI’s d1 models, which Ollama v0.40.12 lists as supported on Apple devices for now. For other Ollama errors, see pull model manifest: 412 and unknown model architecture. For choosing a local model, start with our local LLM for coding guide.