Back to MAX-ALPHA ยท Update history

A more capable MAX-ALPHA

Open capability status and optional smoke checks

Research reviewed October 26, 2026. Discovery dates are not product release dates. Downloadable components are candidates, not automatically installed capabilities.

Implemented: more reliable document excerpts

Edited files now refresh their cached excerpts. Chinese and Japanese phrases can match inside unspaced sentences, and full-width text can match ordinary typed keywords. Source file names, line ranges, bounded context, and omission warnings are retained.

This is lexical retrieval, not semantic understanding or full-document review. It does not read images or scanned pages, recognize every synonym, or increase the model context window. Attached file content remains in the existing local-AI path; this update does not upload files to cloud AI.

What makes an assistant capable

A model is only one component. Reliable assistance also needs source retrieval, bounded memory, isolated tools, explicit permissions, persistent task execution, and tests with real workflows. A download does not transfer dot, Codex, ChatGPT, or Muse accounts, integrations, memory, or capabilities.

Official worldwide options reviewed

Qwen-Agent and Qwen models

The official Chinese Qwen-Agent guide describes tools, planning, retrieval and model-server connections. The framework uses Apache 2.0; individual model licenses must be checked separately. Its code-interpreter example explicitly lacks production sandbox protection.

Decision: do not bolt an unrestricted code interpreter or MCP tools onto this browser. A protected backend, tool allowlist, confirmations and bounded execution are prerequisites. The existing MAX-ALPHA cloud adapter already names Cloudflare-hosted Qwen3.8-27B; changing model names alone would not verify working inference or free-tier eligibility.

llama.cpp

Official source and downloads provide an MIT-licensed runtime with an OpenAI-compatible server. Model weights have their own licenses, memory requirements and quality limits.

Decision: a possible self-hosted cloud backend, not a drop-in browser script. It needs provisioned compute, secured service access and inference benchmarks. No model weights, server, credentials or paid capacity were installed.

Microsoft MarkItDown

Official project converts many office and document formats into Markdown for analysis. It uses the MIT license and a Python runtime. Its security guidance warns that converters inherit process filesystem/network privileges.

Decision: useful candidate for a sandboxed cloud document service with narrow conversion functions, upload consent, size limits and no arbitrary URL access. It is not a browser-native replacement and is not a guarantee of layout-perfect PDF or Excel preservation. No private files were sent to it.

WebLLM

Official WebLLM provides browser WebGPU inference. MAX-ALPHA already pins a WebLLM version and explicit supported models.

Decision: keep the existing optional local route. GPU/device support and actual model tests are needed before upgrading. This route runs on the user's device, so it does not satisfy cloud-only processing. No automatic model download or silent provider switch was introduced.

Mistral

Official releases offer model and assistant tooling candidates. Mistral's license guidance distinguishes Apache-licensed models from models with additional commercial terms.

Decision: compare an exact model and license against the existing backend on representative tasks before switching. Open weights do not include free cloud hosting, paid API access, or another product's Work mode.

Japan: Sarashina

The official Japanese research catalog and Sarashina2.2-TTS project were reviewed. The TTS package uses a noncommercial license.

Decision: do not add this noncommercial package to a potentially commercial app without resolving the license. It is not a general-purpose assistant upgrade.

Next capability gates

  1. Verify the existing hosted service and one real answer through authorized account access. A setup status alone is not inference proof.
  2. Benchmark instruction following, sourced answers, math, coding and multi-step tasks against the current model. Preserve the current route if the candidate is worse.
  3. Before cloud file analysis, provide a separate protected upload/conversion path and explicit data-sharing controls.
  4. Before autonomous jobs or connected tools, add durable task storage, least-privilege tools, cancellation and confirmation gates. Do not treat generated plans as completed actions.

No new paid service, API key, persistent account grant, model download or backend deployment accompanies this browser reliability update. Live inference, cloud runtime and native-device behavior are not established by source tests.