import ChangelogHeader from "@/components/Changelog/ChangelogHeader"

Highlights

Multi-Token Prediction (MTP)

Jan now supports llama.cpp's Multi-Token Prediction for compatible models (e.g. GLM 4.5/4.6 and other architectures with MTP heads). Jan detects MTP-capable models from GGUF metadata ({arch}.nextn_predict_layers) at import time and exposes a per-model toggle plus draft tunables (spec-draft-n-max, spec-draft-n-min, spec-draft-p-min) in Model Settings. When enabled, spec-type = draft-mtp is emitted into the router preset, letting the model draft multiple tokens per step for faster generation. Requires the bundled llama.cpp build b9193 or newer; older backends disable the toggle with an upgrade hint.

llama.cpp Router Mode

Jan's local inference engine now runs as a single unified router process instead of spawning a separate server for every model. The router loads and unloads models on demand, so switching between them is faster and uses less memory. This release also adds:

A reasoning toggle (brain icon in the chat input) to switch thinking mode — auto / on / off — per model without reloading; only available when using the llama.cpp provider
Live prompt-processing progress so you see a percentage while the model processes your input rather than a bare spinner
A "Loading model…" indicator when the router cold-starts a model
A dialog prompting you to confirm or cancel when you close Jan while a model is still working

Inline MCP Tool Approval & Citation Cards

MCP tool calls no longer interrupt the conversation with a blocking dialog. Approval panels now appear inline inside each tool card, showing the exact arguments before you accept or deny. RAG results are displayed as numbered citation cards with source previews inside the tool output, and assistant responses include superscript markers linking back to each matched source (cosine similarity ≥ 0.65). Web search results show citation cards in the tool output but do not inject superscript markers into the response text.

Model Fit Labels & Bulk Delete

The Hub and provider model lists now show a colored fit pill — Fits, May be slow, or Won't fit — based on your hardware, without downloading anything. Model quantizations are grouped as Small, Balanced, or Large with a Recommended tag on the default download. A new Delete All button removes every managed model download at once and shows the total disk space to be freed (imported models are left untouched). Failed downloads no longer get stuck — they are cleared from the queue and a toast is shown. To retry, restart the download from the Hub or provider list.

Provider List Redesign

The provider list in Settings is now split into Local (llama.cpp, MLX) and Remote (OpenAI, Anthropic, Google, Groq, and others) sections so it is immediately clear which providers run on your device and which send data to the cloud.

Per-Thread Activity Indicators & Navigation

The thread sidebar now shows a small spinner next to each thread that is actively generating, loading a model, or running tools. Navigating away from an active thread no longer interrupts it — the work keeps running and the UI restores the correct state when you return. Deleting a thread properly stops all in-flight activity.

Audio Attachments

You can now attach audio files (WAV, MP3) directly in the chat input alongside images and documents. The Add Audio option appears when the active model supports audio input. Attached files show a preview chip with the filename and duration before sending.

File Attachment Progress Bar

File uploads now display a progress bar so you can see how much has transferred before sending your message.

Backend Dependency Checker

After installing a llama.cpp backend, Jan scans for required libraries (CUDA, Vulkan, cuDNN) and displays a checklist of anything missing with links to the official installers.

KV Cache Default Reverted to f16

The default KV cache type was temporarily changed to q8_0 in a prior release and has been reverted to f16 as a safer default. This change is not automatically migrated — if your models fail to load after updating, go to Settings > llama.cpp > KV Cache K Type / V Type and set both back to f16.

Vercel AI SDK v6

OpenAI, Anthropic, and Gemini model lists are now auto-populated when you enter an API key, with capabilities (tools, vision, audio) inferred from model IDs. Fallback API keys persist across restarts. Gemini 3.0 models route through Google's native SDK for correct tool-calling behavior. Error messages for context overflows and template errors are clearer and more actionable.

Server-Side MCP Tool Execution

A new /v1/orchestrations endpoint runs MCP tool calls server-side on localhost:1337, and a Settings toggle lets you enable the same orchestrator on /v1/chat/completions. An optional lightweight router model can be configured to pre-select which MCP servers a request needs, keeping the tool list short for the main model.

Message Queue & Editable Pending Chips

You can now keep typing while a response is streaming. Queued messages appear as editable chips in the input area, send automatically as soon as the current turn finishes, and the Stop button uses a two-stage interaction so you can cancel either the queue or the active generation.

Auto-Summarized Thread Titles

After your first response, Jan summarizes the thread title in the background using a cheap inference pass that can be cancelled if you rename the thread yourself.

Chain of Thought, Regenerate & Edit Improvements

Failed assistant responses now show a regenerate button instead of leaving the conversation stuck. A new Chain of Thought rendering surfaces step-by-step reasoning more cleanly, and you can scroll through streaming thinking content without being auto-interrupted.

Settings, Assistants & UX

Settings navigation has been reorganized
A Default Assistant setting and a chat-bar assistant picker (shown only when you have more than one assistant) make it easier to switch personas
Audio, vision, and document attachment options are now hidden when the active model doesn't support them
Document attachments accept many more file extensions
Toast placement is configurable, and message timestamps render in your local timezone
Hindi, Korean, and Catalan locales were added; existing locales were updated for the new strings
Up Arrow in an empty input restores your last sent message
Factory Reset offers options to preserve user data
All user settings are now stored in local files instead of browser localStorage, preventing data loss on reload

Remote Provider Improvements

MiniMax has been added as a built-in cloud provider, defaulting to M2.7
Configurable token limits and auto-compact behavior for remote providers
Restored Gemini SDK adapter for MCP tool calls, and Groq no longer breaks on follow-up turns after a reasoning response
Mistral Magistral models stream through a dedicated adapter that understands their thinking part shape
Gemma's <thought> reasoning tags are recognized by the parser and rendered as a collapsible reasoning block instead of appearing as raw text in the response

Bug Fixes

llama.cpp context meter now accounts for tool-heavy chats, and ctx_size overflow no longer breaks reloading a chat
CUDA backend is skipped when the GPU has insufficient memory, and CUDA library paths are no longer overwritten by other library setup
Flatpak builds discover CUDA libraries and GL extensions for NVIDIA GPUs
PDF parser panics are caught and a file size limit is enforced on uploads
Stdio MCP transports verify tools/list before signalling ready, and reconnect automatically on disconnect; tool parameter schemas missing type are normalized for strict providers
The MCP startup error dialog is shown again when servers fail to launch
Internal RAG tools used for embedded documents are auto-approved
Project file ingestion progress now persists across navigation, file upload errors are surfaced in plain language, and long URLs no longer overflow the project view
The Hub fit badge and MLX download UI no longer get stuck after a failed download
Spell check setting is honored in the Assistant editor's Instructions and Description fields
Assistant identity preamble is stripped on upgrade from the older v2 default prompt
Jan CLI install detection now works even when the install destination isn't on the app's PATH
App config persists in the Tauri app data directory, and Windows extended-length paths open correctly in File Explorer
Left sidebar maximum width reduced so content no longer cramps
Chat input and new-chat buttons are disabled during engine downloads

Known Issues

KV cache type is no longer auto-migrated. A prior release temporarily defaulted the llama.cpp KV cache to q8_0, which broke loading for some models that require flash attention. v0.8.0 reverts the default to f16 and removes the previous f16 → q8_0 migration code. Existing installs are not migrated automatically — if a model fails to load after updating, go to Settings > llama.cpp > KV Cache K Type / V Type and switch both back to f16 manually.
Migration glitches when upgrading from an older release. If the app behaves oddly after upgrade (settings not applying, model lists looking wrong, etc.), open Settings > General > Factory Reset and run the reset with Keep models and configs enabled. This clears stale app state without deleting your downloaded models or provider configuration.
MLX support is experimental. The MLX engine is still under active development on Apple Silicon and may exhibit instability or compatibility gaps compared to the llama.cpp engine.

Update your Jan or download the latest.

For the complete list of changes, see the GitHub release notes

Jan v0.8.0: Multi-Token Prediction, llama.cpp Router Mode & Inline MCP Approval

Highlights

Known Issues