ollama/ollama
Get up and running with Kimi-K2.6, GLM-5.2, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.
How ollama is put together
Ollama is a tool for running large language models locally, written primarily in Go. It exposes a local REST/NDJSON HTTP API (default port 11434) backed by llama.cpp servers that are discovered, launched, and supervised at runtime; models are pulled from a registry and converted from safetensors/PyTorch formats into GGUF via the convert package. A cobra-based CLI (main.go -> cmd) drives an interactive TUI and an agent framework (agent/) that runs model/tool loops with scoped human approval for dangerous tools. A React/TypeScript desktop chat UI lives under app/ui/app, and native GPU/runtime discovery plus CMake-built llama.cpp runners form the inference layer.
Languages
Frameworks
Datastores
Infrastructure
Major components
cmd
Cobra CLI commands, interactive chat/editor sessions, agent TUI, cloud suggestions, and platform-specific startup logic.
agent
Agentic run loop orchestrating model chat rounds, tool registry/execution, context compaction, skills, and scoped tool-approval state.
api
Typed Go client for the Ollama REST API (/api/generate, /api/chat, /api/pull, /api/push, /api/create) with NDJSON streaming and request signing.
convert
Converts third-party model weights (safetensors, PyTorch pickles, sentencepiece tokenizers) for many architectures into GGUF format.
discover
Detects GPUs and compute backends (CUDA, AMD, Vulkan, Metal on macOS) and manages llama.cpp runner binaries/probes.
llm
Locates, launches, and supervises the llama.cpp server subprocess, including exit-status handling per platform.
auth
Signs outbound requests (used by api client toward ollama.com or when auth is enabled) with a local key.
app/ui/app
React 19 + Vite desktop chat application rendering markdown/math/code with sanitized rehype pipelines.
The new Claude Desktop integration was polished with model mappings, an Auto mode setting, and several responsiveness fixes.
5 weeksWeek of 2026-08-24
Week of 2026-08-24latest
The new Claude Desktop integration was polished with model mappings, an Auto mode setting, and several responsiveness fixes.
Week of 2026-08-17
A brand-new Claude Desktop app with onboarding and model management landed alongside major MLX caching improvements and a metadata cache to speed up every request.
Week of 2026-08-10
New model support (Qwen 3.8, Nemotron 3.5), DeepSeek Harness and Muse Code integrations, web search in the Responses API, and release v0.32.7 all shipped in a busy week.
Week of 2026-08-03
Image input arrived for Qwen 3.5 models on Apple silicon, a DFlash block-diffusion draft model was introduced, and the agent UI gained streamed thinking traces.
Week of 2026-07-27
The DFlash speculative-decoding groundwork was laid out across the MLX runner, GPU kernel helpers were factored out, and lint enforcement was enabled tree-wide.
Over the past five weeks, Ollama's desktop app has grown into a full integration hub, adding a Claude Desktop app with onboarding, model management, and app-connection flows. Under the hood, the team invested heavily in the MLX engine—improving caching, prefill handling, image input, and speculative decoding via a new DFlash draft model—while also shipping new model support (Qwen 3.8, Nemotron 3.5) and OpenAI-compatible API features like web search. Stability work rounded things out, including fixes for streaming errors, data races, and cache corruption.
Week by week
2026-08-24The new Claude Desktop integration was polished with model mappings, an Auto mode setting, and several responsiveness fixes.latest6 changes
Feature
Claude Desktop model mappings
The desktop app now knows how to map requests to the right Claude models.
Feature
Claude Desktop Auto mode setting
Users can toggle an automatic mode for how the Claude Desktop integration behaves.
Fix
Scrollable integrations list
The list of integrations in the app can now scroll instead of being cut off.
Refactor
Desktop integration responsiveness
Interactions with desktop integrations feel snappier after performance improvements.
Fix
Desktop interaction regressions fixed
Bugs that had crept into desktop interactions were repaired.
Fix
Empty editor guard in command line
The command-line tool no longer crashes when opening an empty editor.
2026-08-17A brand-new Claude Desktop app with onboarding and model management landed alongside major MLX caching improvements and a metadata cache to speed up every request.6 changes
Feature
Claude Desktop app added
A dedicated Claude Desktop application shipped with its own onboarding flow and 'Connect your apps' experience.
Feature
Claude model management
The app can now manage which Claude models are available and recommended.
Refactor
Model metadata cache
A new cache cuts down Ollama's per-request overhead by avoiding repeated metadata lookups.
Refactor
MLX prefill snapshot improvements
Cancelled or interrupted prompts no longer lose their cached progress, making resumed responses faster.
Fix
Streaming parser errors no longer wedge chat
A mid-stream parsing error used to stall chat and generate; it now recovers gracefully.
Chore
Engine updates
Both llama.cpp and MLX engines were updated, plus lint cleanups across the codebase.
2026-08-10New model support (Qwen 3.8, Nemotron 3.5), DeepSeek Harness and Muse Code integrations, web search in the Responses API, and release v0.32.7 all shipped in a busy week.6 changes
Feature
Qwen 3.8 support
Qwen 3.8 gained a renderer, developer instructions support, and tolerance for unusual message ordering.
Feature
DeepSeek Harness and Muse Code integrations
Two new coding-tool integrations were registered so they can connect through Ollama's launcher.
Feature
Web search in Responses API
The OpenAI-compatible Responses API can now perform web searches.
Feature
Nemotron 3.5 and Nemotron 3 Nano Omni
Support landed for the Nemotron 3.5 prompt layout and the Nano Omni model on Apple silicon.
Refactor
Faster prefill for nvfp4 models
Prompt processing got faster on double-scale nvfp4 quantized models.
Chore
Release v0.32.7
A new stable release was cut along with routine engine updates and docs fixes.
2026-08-03Image input arrived for Qwen 3.5 models on Apple silicon, a DFlash block-diffusion draft model was introduced, and the agent UI gained streamed thinking traces.6 changes
Feature
Qwen 3.5 image input
Qwen 3.5 models can now accept images, with the runner feeding media features during prompt processing.
Feature
DFlash draft model
A block-diffusion drafting model called DFlash was added to speed up generation.
Feature
Streamed thinking traces in agent TUI
The terminal agent interface now shows the model's reasoning as it streams in.
Fix
Cache rewind corruption fix
Fixed a bug where refilling rewound caches could corrupt later saved snapshots.
Fix
OpenAI-compatible streaming format
Chat completion streams now match OpenAI's exact wire format for better client compatibility.
Fix
Data race fixes
Progress indicators and the macOS build pipeline had race conditions that were cleaned up.
2026-07-27The DFlash speculative-decoding groundwork was laid out across the MLX runner, GPU kernel helpers were factored out, and lint enforcement was enabled tree-wide.6 changes
Feature
DFlash drafting session groundwork
The MLX runner gained the plumbing for block-diffusion drafting, letting models declare their own cache needs.
Refactor
GPU kernel helper scaffolding
Custom GPU kernel code was consolidated into shared helpers to reduce duplication.
Refactor
Recurrent layer optimizations
Gated-delta steps now run in a single launch and convolution outputs are properly activated.
Chore
Lint enabled on whole tree
Code-quality checks now run everywhere, with existing issues cleaned up first.
Refactor
MLX image generation removed
Unused MLX-based image generation code was deleted to simplify the codebase.
Fix
Agent file mentions on Enter
Typing @file mentions in the agent terminal now works when pressing Enter.
Dependencies and code review
Dependency advisories
Security Watch
No known advisories across 0 scanned dependencies.
No known advisories in the scanned dependencies.
Code review
No concrete code-level issues confirmed.
Get this report every week for your repos.
GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.
$19 a month, flat · First 10 outputs free · No card required