v2.0.4
ruixenui/ruixen.comv2.0.4Mar 25, 2026by enricoros
AI Summary
This release focuses on precise hyperparameter configurations for AI models, including reasoning effort, verbosity, and temperature across various providers, alongside support for new frontier models like Opus 4.6 and GPT-5.4.
Key Highlights
- Support for Anthropic Opus 4.6 and Sonnet 4.6 models
- Precise API-validated parameter support (reasoning effort, budget, verbosity)
- New AWS Bedrock integration with native protocols
- Mobile focus mode enhancement
New Features
- 1M token context support
- Configurable web search and fetch depth
- US-based inference option
- Fast mode for Opus (2.5x speed)
- Attachments on any message
- Rich text copy and reasoning trace controls
Full Release Notes
# Big-AGI 2.0.4 - Hyper Params ### Precision matters. This release is about getting models right: accurate parameter ranges for reasoning effort, verbosity, and temperature across providers and through 3rd parties such as OpenRouter and AWS. Plus the latest frontier models, a new cloud platform, and Anthropic going deeper. A huge shout-out goes to @Slava-256, for making Big-AGI's model support the best-in-class with this release. From filing detailed issues for accurate reasoning effort ranges ([#940](https://github.com/enricoros/big-AGI/issues/940), [#944](https://github.com/enricoros/big-AGI/issues/944)), verbosity accuracy ([#947](https://github.com/enricoros/big-AGI/issues/947), [#948](https://github.com/enricoros/big-AGI/issues/948)), temperature support ([#953](https://github.com/enricoros/big-AGI/issues/953)), to pushing for Opus 4.6 support ([#962](https://github.com/enricoros/big-AGI/issues/962)), and [many more (31 closed tickets!)](https://github.com/enricoros/big-AGI/issues?q=state%3Aclosed%20author%3ASlava-256). Slava's thorough testing, precise reports, findings, and reference code are the reason why 2.0.4 was focused on precise parameters support. <img width="1792" height="1024" alt="0 - A2 - High-key_white_scene_with_impeccable_cleanliness__employing_heroic_framing _A_detailed_close-up_view_of_a_capybara__compellingly_crafted_from_crystal_glass _The_capybara_poses_with_flair__sporting_sleek_black_sunglasses_and_securely_positio" src="https://github.com/user-attachments/assets/257df06f-d365-4bf6-a81a-a9e74f97c0fe" /> ### Models & Parameters - Anthropic **Opus 4.6** with adaptive thinking and 1M tokens, **Sonnet 4.6** and new features - Google **Gemini 3.1 Pro**, **Nano Banana 2** - OpenAI **GPT-5.4** family with 1M context and reasoning controls - xAI **Grok 4.20** - New **Z.ai** GLM models support, with configurable thinking - **API-validated parameters**: reasoning effort, budget, verbosity, temperature, tools, are tested on live APIs - **Model-specific controls**: each model exposes only the parameter values it actually supports - **Automated sweep testing** across native APIs keeps parameter definitions current ### Anthropic API mastery - Configurable web search and web fetch depth, and dynamic search - API controls such as **US-based inference** and **Fast mode for Opus** (2.5x speed, at 6x cost) - Continuation for long outputs (pause_turn support) ### AWS Bedrock support (New) - Authentication via both IAM credentials, or using a single AWS Bedrock token - Native support for 3 native AWS Bedrock protocols: Anthropic models via Invoke, Amazon Nova models via Converse, and other models via OpenAI-compatible ChatCompletions ### Also New - Attachments on any message, lossless images option (warning: can be very large on Gemini models) - Rich text copy, reasoning trace controls - System theme mode follow - [Mobile focus mode](https://github.com/enricoros/big-AGI/issues/894) ## New Contributors * @Blacksuan19 made their first contribution (SSR guard, system theme mode) in https://github.com/enricoros/big-AGI/pull/1011 * @dLo999 made their first contribution (local time for backup exports) in https://github.com/enricoros/big-AGI/pull/1041 * **Full Changelog**: https://github.com/enricoros/big-AGI/compare/v2.0.3...v2.0.4 ## Get Started Available now at [big-agi.com](https://big-agi.com), via Docker, or self-host from source.