altxai

Next Task: Awaiting DB & WebGPU Init

Idle 0%

Preparing engine… Estimating time remaining…

System Process Logs

System initialized. Storage connected to IndexedDB.

Upload files to inject them into the active workspace database. Open settings to pick an Altx model and load the local inference engine.

v2.8.0 - Connectivity Pre-Flight Check

Added a quick connectivity check to huggingface.co before starting a download, so a hard network block is reported immediately and accurately instead of waiting through 4 retry attempts to find out.
Widened network-error detection to also catch plain "Failed to fetch" (not just Cache.add() errors), since both mean the same underlying thing: the request never completed.
If basic connectivity succeeds but the actual model download still fails on every retry/backend, the error message now says so explicitly — that pattern usually points to something blocking large CDN downloads specifically (ad/privacy blocker, VPN, corporate firewall) rather than a totally offline connection.

v2.9.0 - Hardened Download Engine

Pinned @mlc-ai/web-llm to v0.2.84 — the unpinned CDN import could silently load a breaking version on any page reload.
Increased download retries from 4 to 8 attempts with exponential backoff (1s, 2s, 4s … capped at 20s) instead of flat 800ms delays.
On attempt 5, the app now nukes any partially-downloaded shards before retrying — corrupted/incomplete shards from a previous session no longer cause every future attempt to fail with the same error.
Cache strategy now alternates every attempt (Cache API → IndexedDB → Cache API …) instead of switching once at attempt 3.
Widened network error detection to also catch AbortError, net::ERR_* Chromium codes, and Safari's "Load failed" — these were falling through to the generic GPU error path before.
Added HuggingFace CDN probe (cdn-lfs.huggingface.co) alongside the main site probe — the CDN can be blocked even when the homepage loads.
Added a disk space pre-flight check using the StorageManager API — warns before starting if free space is below what the selected tier needs (~1.2GB Lite, ~2.8GB Eco, ~5.2GB Pro).
Added a WebGPU adapter check at boot — if the browser supports WebGPU but no GPU adapter is found (VM, missing drivers, remote desktop), the error is reported immediately with a clear explanation instead of failing deep inside the engine load.
Error messages now correctly distinguish GPU/VRAM failures from network failures, and suggest a smaller tier when relevant.

Renamed all tiers to Altx Lite, Altx Eco, and Altx Pro. The model dropdown now only ever lists these three — no raw bundled-model fallback section.
Added an Experimental Features section in Settings: Thinking and Experimental Mode are both off by default and must be turned on manually.
With Thinking enabled, the model reasons in a separate collapsible panel before writing its answer, instead of mixing reasoning into the visible reply.
Reasoning is now capped to a short budget so the model commits to an answer instead of thinking indefinitely.
File edits now render as a live red/green diff (additions and removals), shown in a slightly muted gray panel separate from normal chat text.
File writes show a dedicated card with filename and status, rather than streaming raw file contents directly into the chat bubble.
Redesigned the model-loading UI with a custom animated loader and an estimated time remaining readout.
Added a "What Altx Can Do" feature list to Settings covering line edits, refactors, debugging, tests, docs, and more.

v2.6.0 - Switched Alt Tiers to Stable Qwen3

Alt Lite, Alt Eco, and Alt Pro now load real, MLC-bundled Qwen3 models (1.7B / 4B / 8B, q4f16_1) instead of Qwen3.5.
Qwen3.5 tiers were removed: that architecture (Gated DeltaNet + MoE) isn't supported by the published WebLLM runtime yet, which is why those tiers would compile-load forever and never finish.
Tier verification now checks the bundled webllm.prebuiltAppConfig directly instead of fetching configs from Hugging Face, so loading is faster and works offline-from-HF.
Download progress still shows real percentage and stage text for whichever tier you load.

v2.5.0 - Qwen3.5 Tiers & Live Verification

Added three named tiers — Alt Lite (0.8B), Alt Eco (4B), Alt Pro (9B) — backed by the official mlc-ai Qwen3.5 MLC builds on Hugging Face.
The app now checks each model's mlc-chat-config.json on Hugging Face live before listing it as ready, instead of trusting a hardcoded assumption.
Known-stable Qwen3 builds remain selectable as a fallback tier, since Qwen3.5's Gated DeltaNet + MoE architecture is newer and may not run on every browser/WASM combination yet.
Download progress now shows real percentage and stage text for whichever tier you load.

v2.4.0 - Native Icons & Live Model Sync

Replaced the Material Icons web font with inline SVG icons, so the UI no longer breaks or misaligns when icon fonts fail to load.
Added a real download progress bar with percentage and stage label during model compilation.

v2.3.1 - Storage Layer Restored

Rebuilt the IndexedDB persistence layer (init, save, list, clear).
Fixed a corrupted task-log handler that broke app boot.
Workspace files are now listed and individually removable in Settings.