github.com/ollama/ollama

Clef decision models

Ollama now supports Clef and Clef Flash, Cloudflare's new open-source decision models, through /v1/systemone.

Clef (27B) and Clef Flash (9B) are multimodal: requests can now include images alongside the text state, shared by all questions and scored jointly with it.

curl http://localhost:11434/v1/systemone -d '{
  "model": "clef-flash",
  "state": "The user took this screenshot.",
  "images": ["<base64-encoded image>"],
  "questions": {
    "has_ollama": {"type": "noul", "instructions": "Does this image contain Ollama?"}
  }
}'

{
  "model": "clef-flash",
  "answers": {
    "has_ollama": {
      "type": "noul",
      "noul": 0.958
    }
  },
  "usage": {
    "input_tokens": 548,
    "output_tokens": 0
  }
}

What's Changed

• Models using web search can now perform up to ten searches per response, up from three
• Modelfiles now support CAPABILITY declarations, so model creators can explicitly declare what a model can do. Declarations are preserved when creating from GGUF or safetensors, through model inheritance, and on Modelfile export
• ollama show and the model list now report only decision as the capability for decision models, so clients no longer offer them for general chat, tools, or thinking
• Updated llama.cpp and the MLX engine

Full Changelog: v0.35.0...v0.35.1

ci: fix missing build context (#18742)

models: add clef support (#18741)

Add CAPABILITY declarations to Modelfiles and an additive capabilities
field to create requests. Preserve declarations across GGUF and safetensors
creation, inheritance, and Modelfile export.

Require decision capability before scheduling System One requests instead
of matching Qwen architecture/renderer metadata. Retain main's GGUF-only
scoring restriction until the separate MLX runtime work lands.

Extracted from the capability foundation in 36d46a0 on system_one_mlx;
MLX scoring and manifest-list changes are intentionally separate.

Decision models

Ollama now supports decision models through /v1/systemone, based on TypeSafe’s Jev API.

Decision models return choices, probabilities, and scores instead of text. Use them for tasks such as ticket triage, model routing, and content classification.

Available models:

• Nimble from Bespoke Labs
• Tev1 from Together AI

ollama pull nimble

Send context and one or more questions:

curl http://localhost:11434/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "nimble",
    "state": "Our checkout has returned 500 errors since 9am.",
    "questions": {
      "label": {
        "type": "choice",
        "instructions": "Which label fits this ticket?",
        "criteria": {
          "billing": "Payments and refunds",
          "bug": "Software errors",
          "account": "Login and account access"
        }
      }
    }
  }'

Example response:

{
  "model": "nimble",
  "answers": {
    "label": {
      "type": "choice",
      "choice": "bug",
      "probabilities": {
        "billing": 0.0125,
        "bug": 0.9781,
        "account": 0.0093
      },
      "confidence": 0.8906
    }
  },
  "usage": {
    "input_tokens": 174,
    "output_tokens": 1
  }
}

The API supports three question types:

• choice: Select an option and return probabilities for each.
• noul: Return the probability that a condition is true.
• score: Return a score across an ordered set of criteria

What's Changed

• Settings now opens without waiting for model discovery.
• Fixed the macOS update menu and icon not reflecting an available update at startup.
• Fixed stalled MLX model downloads hanging indefinitely.
• Requests containing the deprecated typical_p parameter now log a warning instead of failing.

Full Changelog: v0.34.4...v0.35.0

mlx: bound pull stall retries and let the watchdog interrupt them (#1…

feat: add System One scoring API (#18606)

What's Changed

• Structured outputs on thinking models now apply in a single pass, making them faster and more reliable.
• Fixed intermittent "model not found" errors with a large local library
• Fixed the macOS app becoming unresponsive when checking if ChatGPT or Codex is running.
• Qwen 3.8 prompt processing is faster on Apple Silicon.
• Gemma 4 on Apple Silicon now picks the best image resolution per image, keeping more detail in high-resolution images.
• Updated llama.cpp, MLX, and XGrammar.

Full Changelog: v0.34.3...v0.34.4

We pick up schema fixes for typed dictionary values and short arrays.

• mlx: speed up Qwen 3.8 prompt processing

Use MLX's gated-delta kernel for long scans and fold dense MLP global scales into SwiGLU.

• address comments

What's Changed

GET /api/show now advertises each model's thinking controls and default:

Available in the CLI with:

ollama show gemma4

    thinking
        levels     false, true
        default    true

Available in the API with:

curl http://localhost:11434/api/show -d '{"model": "glm-5.3-flash:cloud"}'

{
  "thinking": {
    "values": ["low", "high", "max"],
    "default": "max"
  }
}

Also available on ollama.com directly for cloud models.

• Nemotron H vision models are now supported on Apple Silicon with MLX
• Ollama's macOS app will now no longer reopen windows you've closed when activating it
• Fix for model pulls from HuggingFace

Full Changelog: v0.34.2...v0.34.3

server: allow registry cross-host redirects among allowlisted hosts (…

api: expose model thinking levels and defaults (#18473)

What's Changed

• Added first-run setup when running ollama, with options to sign in or continue locally. Setup completion is shared with the desktop app on macOS and Windows.
• Added ollama://apps to open the desktop app’s Apps page directly on macOS and Windows.
• Fixed excessive memory growth during long generations with MLX speculative decoding.
• Updated llama.cpp.

Full Changelog: v0.34.1...v0.34.2

cli: add first-run onboarding shared with the desktop app (#18495)

The decode loop releases MLX's pool of freed buffers every 256 generated
tokens, which is also how often the KV cache grows and drops its previous,
smaller buffers. The check fires only when the token count lands exactly on
a multiple of 256. Speculative decoding emits several tokens per round, so
most rounds step over the boundary and the pool is never released. Each
growth at a long context leaves several GB of buffers that no later
allocation can reuse, so the runner's footprint keeps climbing over a long
generation until the system runs out of memory.

We now release the pool whenever a round crosses a multiple of 256 tokens,
which is what a single-token round already did. With qwen3.8:27b-mlx at a
98k-token context on a 128 GB machine, a long speculative generation
previously grew the runner past 90 GB and panicked the kernel; it now stays
flat at 30 GB.

model is one package with three jobs: the contract between the runner
and the architectures, the opened checkpoint, and building nn layers
from checkpoint tensors. Its files did not say which was which. base.go
carried the folded package's name over the interfaces and the registry,
root.go held the safetensors header scan next to Root, and quant.go
mixed the nvfp4 global-scale helpers with quant parameter resolution.

base.go becomes model.go, named for what it holds. root.go keeps Root
and Open; TensorQuantInfo and the header scan join quant.go, so
everything the checkpoint says about quantization is read and resolved
in one file. The global-scale helpers move to globalscale.go with their
tests. Root.Close, a no-op with one caller, goes. No code changes
otherwise.

llama.cpp build changes resulted in duplicate symbols between libllama and libmtmd. This moves the compat patch into libllama with exported symbols.

What's Changed

• MLX safetensors ollama create no longer experimental. GGUF model creation now requires using llama.cpp tooling for safetensor conversion and quantization.
• Improved MLX memory handling on Apple Silicon
• Runaway repeat token detection now requires 100 repeat tokens for reduced false positives (e.g. OCR)
• /api/tags is much faster on large model libraries (3.1 s → 294 ms cold in testing), and model capabilities are now reported consistently.
• Deprecated typical_p: it can no longer be set when creating new models, existing GGUF models retain support.
• MLX and llama.cpp updates

Full Changelog: v0.34.0...v0.34.1-rc1

Drop support for creating new models with typical_p parameters, while
retaining support for existing GGUF models with the setting.

mlx: add mlx patch to docker build context (#18440)

• MLX: version bump

• mlx: support ModelOpt global scales in MoE models

• address comments

• address comments

Use Ollama models in ChatGPT Desktop

Ollama models can now be used directly in ChatGPT Desktop, so you can keep your existing workflow while running open models. Setup is available from the Ollama app on MacOS.

This release also improves structured output performance on Apple Silicon, adds support for OpenAI-compatible client tool search and response compaction.

Full Changelog: v0.33.3...v0.34.0

openai: support standalone named function outputs (#18348)

proxy: normalize namespaced commands in Full Access (#18331)

openai: accept plaintext-labeled Codex agent messages (#18329)

openai: finalize responses at the web search limit (#18328)

app: harden Codex desktop proxy handling (#18244)

app: add Ollama to ChatGPT Desktop (#18236)

What's Changed

• gemma4 now supports images and audio on MLX engine
• Report cached prompt tokens
• Honor GGUF model defined default parameters
• MLX, MLX-C, llama.cpp update

New Contributors

• @marcelpetrick made their first contribution in #17579

Full Changelog: v0.33.2...v0.33.3

Safetensors gemma4 imports served by the MLX engine now answer image
and audio chats. Images run through both vision architectures: the
transformer tower (26B, 31B, e-series) and the 12B's encoder-free
unified embedder. Audio arrives through the same intake the ollama
API already accepts for gemma4 GGUFs — WAV bytes in the images field,
OpenAI input_audio parts, and /v1/audio/transcriptions uploads — with
the e2b/e4b checkpoints running clips through their conformer audio
encoder and the 12b unified checkpoint embedding the raw waveform
directly. Clips longer than 30 seconds are split evenly into chunks
of at most 30 seconds, cut at pauses, and encoded independently.

Each modality serves only checkpoints that carry it: 26B/31B have no
audio config and reject audio input, and checkpoints with an
unrecognized vision architecture still load as text-only models and
reject image requests.

The server previously hid the vision and audio capabilities for
gemma4 safetensors because the engine served neither. Both
suppressions are removed, and existing imports start advertising the
capabilities without re-importing since import already records them.

llama.cpp: version bump b10760 (#18199)