v2.1.0

# Big-AGI 2.1.0 - Weights Dust

### Six new vendors.

Cerebras, NVIDIA NIM, Sakana AI, Cohere, Modular and Nous Research, each with a setup guide. Plus model updates across Anthropic, OpenAI, Gemini, xAI, Moonshot and Z.ai, a multi-vendor transcription framework, explicit reasoning-effort and mode controls, and code sandboxes that persist across turns. 628 commits since 2.0.5.

<img width="1536" height="1024" alt="210-d4" src="https://github.com/enricoros/big-AGI/releases/download/v2.1.0/210-d4-FINAL.png">


### New Vendors
- **Cerebras** first-class vendor support, with live catalog and CSF (#1146) - [setup](https://big-agi.com/docs/connect-cerebras)
- **NVIDIA NIM** with a curated catalog and sync tooling - [setup](https://big-agi.com/docs/connect-nvidia-nim)
- **Sakana AI** on the Responses API: Fugu Cyber, Fugu Ultra 1.1, Namazu (#1142) - [setup](https://big-agi.com/docs/connect-sakana)
- **Cohere** model family, including Aya - [setup](https://big-agi.com/docs/connect-cohere)
- **Modular** with MiniMax M3, Kimi K2.7 Code and Gemma 4 - [setup](https://big-agi.com/docs/connect-modular)
- **Nous Research** Hermes models and the Portal subscription gateway, and **EmpirioLabs** (#1129), both via [custom endpoints](https://big-agi.com/docs/connect-custom-endpoints)

### Models
- Anthropic **Opus 5**, **Sonnet 5**, **Opus 4.8**, plus **Fable 5** and **Mythos 5**
- OpenAI **GPT-5.6** Sol/Terra/Luna, with Reasoning Mode (#1158) and Max effort (#1159)
- Google **Gemini 3.5 / 3.6 / 3.7 Flash**, **Nano Banana 2 Lite**, Robotics-ER 2, and 3.7 video understanding
- xAI **Grok 4.5**, **4.6** and **Grok Build 0.1**, with none/x-high efforts (#1104)
- Moonshot **Kimi K3**, and Kimi Code subscription keys (#1161)
- Z.ai **GLM-5.2 / 5.3**, Alibaba **Qwen 3.7 Flash**, **DeepSeek V4** Flash and Pro, **MiniMax M3**

### Transcription (New)
- Multi-vendor speech-to-text framework, with its own Settings section - [voice docs](https://big-agi.com/docs/feature-voice)
- **Deepgram** diarization, dictionaries, topic and sentiment detection, word-level spans
- **OpenAI** gpt-transcribe with keyword and language hints, and diarization
- Per-engine instances and profiles, sharing the config layout with Speech and Draw

### Sandboxes & Containers
- Anthropic: standalone **Code Sandbox** toggle, container re-attached across dynamic-web turns (#1087)
- OpenAI: **container files** as hosted resources - download, embed, sandbox-link interception; code execution downgrades to function calls when a container expires
- Gemini: **Omni**, video output, Antigravity agent events, and the Interactions protocol rewrite with usage metrics (#1143)

### Reasoning & Caching
- OpenAI Responses: reasoning phase (commentary vs final) and client-controlled policy, defaulting to all-turns on GPT-5.4+
- OpenRouter: reasoning on/off/default (#1138), Anthropic-style prompt caching via ephemeral breakpoints, session affinity
- Anthropic: reasoning token accounting, x-high auto-probing, cache breakpoints capped at the API limit (#623)
- Vertex AI redirect links resolved automatically or on demand (#1114)

### Reliability
- Multi-tab rewritten on Web Locks; demote-by-wipe replaces the old heuristic (#802)
- Warning that chats live in the browser and are lost on cache clear (#672)
- Auto-reload once on stale-deploy chunk load errors, and denoised error reporting
- Browser support floor enforced in lint, Intl.Segmenter fallback, tiktoken preload failures tolerated

### Also New
- **Draw**: OpenRouter image generation on the dedicated Image API, engine-instance settings
- Settings redesign, Diagnostics section, model assignments with per-domain Auto picks
- Token stats during generation, not just at the end (#1125)
- Markdown preview for .md fenced code (#1126), rendered ```svg blocks, Live-SVG
- Mobile wake lock, model-select search that keeps focus, attachment rename
- Removed: Helicone (maintenance mode), OpenPipe (discontinued), react-player

### Developers
- **AIX Protocol Lab**: protocol microscope and differential tester for the decode layer, alongside the [AI Inspector](https://big-agi.com/docs/feature-ai-inspector)
- Anthropic and Claude Code protocol eviscerators, standalone storage and diagnostics tools
- Per-vendor model-defs manifest with selective boot refresh: model updates no longer need an AIX roll
- Tests wired into CI, type-checking extended over tools/

## New Contributors
* @Adam-Dalloul made their first contribution in https://github.com/enricoros/big-AGI/pull/1130
* @mahirhir made their first contribution in https://github.com/enricoros/big-AGI/pull/1145
* @Robinnnnn made their first contribution in https://github.com/enricoros/big-AGI/pull/1156

**Full Changelog**: https://github.com/enricoros/big-AGI/compare/v2.0.5...v2.1.0

## Get Started
Available now at [big-agi.com](https://big-agi.com), via Docker, or self-host from source.