Big-AGI 2.1.2: Render Tree

Edit anything, run the newest models, and a security fix you should not skip.

Two months of work since 2.1.1: the chat message becomes fully editable as a session, the model catalog jumps a generation, Meta AI joins as a vendor, and a privately reported key disclosure is closed. Self-hosters with a server-side Gemini key should upgrade and rotate it.

Security

• Server-side GEMINI_API_KEY could be read through the AI Inspector request echo on deployments that set it; fixed, see the advisory. Rotate the key after upgrading.
• Next.js 15.5.28 (closes CVE-2025-29927 and the other open 15.1 advisories), Node 24.22.1 base image (CVE-2026-111497, the Brotli decoder fix)
• Opt-in HTTP Basic Auth middleware hardened: passwords with colons, UTF-8, constant-time comparison
• SECURITY.md with a private reporting channel

Models & Vendors

• OpenAI: GPT-6 Astra, Sol and Luna, GPT-6.1 Sol, service tiers (Ultrafast, Fast, Flex), GPT Image 2.5
• Anthropic: Opus 5.5, Fable 5.1 and Mythos 5.1, Sonnet 5.5, Haiku 5.5; pause_turn continuations with stats, automatic caching with hosted tools, Files discard policy
• Google: Gemini 4 Argon, Gemini 3.8 Flash and Flash-Lite TTS, Nano Banana 2.1 with editing, URL Context separate from Search, Antigravity 09-2026
• Meta AI (Muse) as a new vendor over the Responses protocol, including Muse Image
• DeepSeek V4.1 Flash, Grok 4.7, Kimi K3.1 Pro, GLM-5.3, Qwen 3.8 Max, Mistral Large 4, NVIDIA Nemotron 3.5; BaseTen through Custom OpenAI
• Models changelog and Update All (#1204), typed parameter tables (#1228), per-call web search fees and cache-write pricing

Chat

• Edit session: reorder blocks, edit documents, add text and attachments, confirm on conflicting edits
• Attachments placement at a {{attachments}} line, message order preserved (#1234, #1242)
• Message timestamps, markdown in user messages (#1230), confirmed deletions, folder assignment from the side panel, archived chats in search
• Render tree: adaptive streaming markdown (#1086), HTML blocks at message end (#596), diagrams, task lists, Chart.js, user text collapser
• Hosted files: one chip and menu for Anthropic, OpenAI and Gemini files (#1207, #1208)

Reliability

• Retries honor Retry-After within a budget, in-band OpenRouter and OpenAI Responses retries, 429 denylist by upstream code (#1210)
• Server keys never go to a client-set host (Anthropic, OpenAI-compatible, Gemini)
• Beam: merges wait for generating replies and stream into the output, empty and error-only replies are not merge inputs (#1212, #1216, #1196)

Also New

• Export chats for Claude Code (resumable session, chat folder zip)
• Call: configurable end-of-turn pause (#1188)
• Transcription: Gemini 3.5 Transcribe; Speech: Gemini TTS engine
• Browse: bounded request count, screenshot size and HTML cleanup cost
• Developers: AIX protocol lab chain fidelity, model fingerprint harness, icon-barrel lint guard

New Contributors

• @qirh made their first contribution (Perplexity integration attribution header) in https://github.com/enricoros/big-AGI/pull/1229
• @johnfelipe made their first contributions (Browse input bounds, cleanHtml ReDoS, Bedrock region validation) in https://github.com/enricoros/big-AGI/pull/1250, https://github.com/enricoros/big-AGI/pull/1251, https://github.com/enricoros/big-AGI/pull/1252 Full Changelog: https://github.com/enricoros/big-AGI/compare/v2.1.1...v2.1.2

Get Started

Available now at big-agi.com, via Docker, or self-host from source.