What's Changed Qwen3.5 is faster on Apple GPUs: the MLX engine now uses the model's MTP head for speculative decoding automatically /v1/chat/completions streaming now matches OpenAI's wire format: role only on the first chunk, finish_reason on its own chunk, and usage in a separate chunk with stream_options.include_usage. Truncated OpenAI responses now report finish_reason: "length" instead of "to…
llama.cpp update (#17545)
What's Changed Fixed an MLX Metal bug that could reduce output quality for NVFP4 models, particularly Laguna. Full Changelog: v0.32.4...v0.32.5
mlx update (#17397)
What's Changed Support Laguna on Apple GPUs via the MLX engine Quantize draft-model output heads at the requested type when creating speculative-decoding drafts. Fixed Qwen3 MoE decoding for differently-quantized experts, plus faster packed gate/up projection (~4–9% on M5 Max). Full Changelog: v0.32.3...v0.32.4
model: add Laguna MLX support Add Laguna XS 2, XS 2.1, and S 2.1 support to the MLX model and create paths. Read the source config to apply one quantization policy across dense and routed MoE layers. Keep the tied output head and router at source precision, quantize supported attention and expert projections, selectively promote sensitive expert down projections, and emit per-tensor metadata for m…
Withdrawn, please use 0.32.3 or newer
What's Changed Fixed model downloads that stall before sending data. Improved integrations: restored Claude Code Channels, fixed Anthropic thinking streams, and made Hermes Desktop respect --force-build. Expanded GPU support with CUDA on Windows ARM64, B200 support through CUDA 12, and lower memory use on Linux CUDA/ROCm iGPUs. Added chat, thinking, and tool calling support for Laguna 2.1 models,…
Update llama.cpp to pick up upstream Laguna implementation and remove Ollama's local Laguna implementation. Retain a narrow Metal-only scaling workaround for routed-MoE prompt overflow. Translate older Ollama GGUF attention-gate and SWA metadata names so existing models continue to load.
This refactors the existing integration tests into 3 priumary groups: fast, release, and library. It also refines some of the release tests to drop some of the older models and pick up newer models, while retaining the broad coverage in the library group.
Needed for cross-compiling WoA
server: detect download stalls before the first byte server: keep stall timeout out of download API
Add compute capability 10.0 to the Linux CUDA v12 preset so B200-class devices can use the cuda_v12 backend with drivers that do not meet the CUDA v13 minimum. Fixes #12583
What's Changed Improved Gemma 4 tool calling and multi-turn reasoning, including more reliable tool-response continuations Fixed a recurrent MLX model cache leak that could increase memory use across requests, and improved cache snapshot performance MLX text model loading now respects OLLAMA_LOAD_TIMEOUT Agent web search and fetch now tell users to run ollama signin when authentication is required…
cmd: put current working dir in the system prompt (#17188)
What's Changed New interactive agent experience: running ollama now launches an agent to help you code and delegate work ❯ ollama Ollama 0.32.0 ▸ Chat, Code, & Work (glm-5.2:cloud) Chat with models, code, search the web, and delegate real work Renamed the Codex App integration to ChatGPT: use ollama launch chatgpt (and --restore to return to your usual ChatGPT profile) Simplified integration selec…
cmd: agent UI (#17017)
What's Changed Enabled flash attention on older NVIDIA GPUs (compute capability 6.x) iGPU can now offload vision models with padding to fit available memory Fixed structured output for thinking models when thinking is disabled Hardened GGUF model creation ollama launch for Claude Code now disables telemetry by default Fixed loading models on paths with non-UTF-8 characters Updated the MLX and llam…
llm: allow iGPU mmproj offload with fit padding llama.cpp's fit pass sizes text-model placement before the multimodal projector is loaded. Ollama had been avoiding that risk on non-Metal iGPUs by disabling projector offload entirely, which forces CLIP onto CPU on GB10 and Strix Halo even when the projector has ample memory available. Let integrated GPUs use the same projector-memory check as other…
create: harden GGUF create flows lint
mlx: update to de7b4ed9 (#17056)
Faster Gemma 4 on Apple Silicon Gemma 4 is now significantly faster in Ollama on Apple Silicon, generating tokens nearly 90% faster on average across a coding-agent benchmark by leveraging multi-token prediction (MTP). Ollama auto-tunes how many tokens to draft as it runs, so the speedup is on by default, requires no configuration, and does not change the model's output. What's Changed Tightened G…
launch: check for min version for hermes desktop (#16912)
What's Changed tools: ignore braces inside JSON strings when detecting tool call end by @aditya-786 in #16937 mlx: bump dependency by @dhiltgen in #16935 llama.cpp update by @dhiltgen in #16960 New Contributors @aditya-786 made their first contribution in #16937 Full Changelog: v0.30.11...v0.30.12-rc0
What's Changed launch: add thinking capability detection to opencode by @hoyyeva in #15434 launch: auto-install Claude Code by @hoyyeva in #16802 launch: auto-install opencode when missing by @hoyyeva in #16806 discover: fix inverted iGPU/dGPU Vulkan classification on Windows hybrid graphics by @Sahil170595 in #16669 mlxrunner: unify and tune speculative decoding by @jessegross in #16791 launch/co…
parser/renderer: add Ornith 9B renderer/parser support (#16920)
llama.cpp version update (#16548)
What's Changed Command A and North family models now run on Apple Silicon with the MLX engine Updated the underlying llama.cpp engine to build 9672 Fixed build artifacts for MLX Full Changelog: v0.30.9...v0.30.10
ci: pin darwin release xcode (#16788)
llama: update llama.cpp to b9672 (#16775)