⚡ Performance Improvements

• Outlines Core: Enjoy faster FSM index construction with a new implementation (#1175).
• 98% Reduction in Runtime Overhead: Reduced overhead by storing FSM-token-mask as tensors. (#1013)

🚀 New Features

• Transformers Vision Models: Apply structured generation with vision + text inputs (#1052)
• OpenAI-Compatible API Support: Use models.openai with any OpenAI-like API (e.g. vLLM, ChatGPT), including structured generation with generate.json and generate.choice (#1142).

💡 Enhancements

• Unified Logits Processors: All models now use shared outlines.processors, completed by adding the following to the integration: llama-cpp, vLLM and ExLlamaV2).
• Custom Regex Parsers: Simplify the implementation of custom Guide classes with Regex Parser support (#1039).
• Qwen-style Byte Tokenizer Support: Now compatible with Qwen-style byte tokenizers (#1153).

🐛 Bug Fixes

• CFG Beta: Fixed large number of bugs to enable beta version grammar-based generation using Lark (#1067)
• Fixed incorrect argument order breaking some models in models.transformers_vision (#1077).
• Resolved OpenAI fallback tokenizer issue (#1046).
• Option to disable tqdm bars during inference with vLLM (#1004).
• models.llamacpp no longer includes implicit max_tokens (#996).
• Fixed whitespace handling for models.mlxlm (#1003).
• models.mamba now working, and supporting structured generation (#1040).
• Resolved pad_token_id reset issue in TransformerTokenizer (#1068).
• Fixed outlines.generate generator reuse causing runtime errors (#1160).

⚠️ Breaking Changes

• outlines.integrations is now deprecated: #1061

Full Changeset

• Add contributors, creation and update date by @rlouf in #1000
• Add SimToM prompt recipe by @kstathou in #1002
• Ensure models.llamacpp Doesn't Have Implicit max_tokens by @lapp0 in #996
• fix models.mlxlm whitespace prefix handling by @lapp0 in #1003
• Adding the option to avoid displaying tqdm bars at inference with vllm by @BIMM99 in #1004
• Update pyproject.toml, enable mlx-lm requirement only on darwin, disable vllm requirement on darwin by @lapp0 in #1005
• Add asknews example by @robcaulk in #1008
• Improve outlines.processors, add integration tests to test_generate.py by @lapp0 in #998
• Abridged version of the .txt article on Coding For Structured Generat… by @willkurt in #1012
• RegexFSM: Cache Legal-Token Mask as torch.Tensor to Improve Performance by @lapp0 in #1013
• JSON response format fix by @rshah713 in #1029
• Fix broken link in README.md regarding Serving with vLLM by @shinohara-rin in #1030
• Major bug & fix: Fix bug in batched multi sample generation by @JulesGM in #1025
• Update file contributors style for docs by @gtsiolis in #1034
• fix: ocd missing blank space by @talboren in #1036
• Adds support for custom regex parsers (for multimodal structured generation) by @leloykun in #1039
• Update models.transformers to use SequenceGeneratorAdapter and OutlinesLogitsProcessors by @lapp0 in #966
• Use outlines.processors for models.llamacpp by @lapp0 in #997
• Add QA with Citations example to the Cookbook by @alonsosilvaallende in #1042
• Fix mamba integration by making it a variant of outlines.models.transformers by @lapp0 in #1040
• fix PyPI name for autoawq by @davanstrien in #1045
• add fallback tokenizer by @JerryKwan in #1046
• Update cerebrium instructions by @milo157 in #1047
• Add Knowledge Graph Extraction example to the Cookbook by @alonsosilvaallende in #1049
• Correct link and add llama-cpp-python installation instructions by @alonsosilvaallende in #1051
• Introduce outlines.models.transformers_vision by @lapp0 in #1052
• Use outlines.processors and SequenceGeneratorAdapter for outlines.models.vllm by @lapp0 in #1053
• Modal documentation: fix deprecated memory= parameter by @Perdu in #1057
• Documentation: Fix failing Modal example by @Perdu in #1058
• Add links to the two examples by @alonsosilvaallende in #1062
• Add link to Multimodal Structured Generation (MMSG) library to docs by @leloykun in #1064
• Correct the documentation to disable caching by @alonsosilvaallende in #1069
• Fix TransformerTokenizer pad_token_id reset by @ispobock in #1068
• Fix link to mamba model reference by @rlouf in #1072
• More detailed mlxlm documentation by @lapp0 in #1074
• Add chain of thought example by @alonsosilvaallende in #1087
• Fix coverage exclude_lines setting for ellipsis by @brandonwillard in #1089
• Add ReAct agent example by @alonsosilvaallende in #1090
• Include SPDX license info in project metadata by @tiran in #1094
• Update Documentation by @lapp0 in #1063
• Make model_class required arg, default processor_class to AutoProcessor by @parkervg in #1077
• Fix details in the documentation by @rlouf in #1096
• Change cookbook examples: Download model weights in the hub cache folder by @alonsosilvaallende in #1097
• Correct variable name in chain-of-thought example by @cpfiffer in #1101
• Remove deprecated outlines.integrations by @rlouf in #1061
• Use relative coverage source paths by @brandonwillard in #1113
• Add missing CI matrix step by @brandonwillard in #1124
• Update modal example by @cpfiffer in #1111
• docs: fix typo by @billmetangmo in #1120
• Pass text and images as kwargs to VLM processor by @lapp0 in #1126
• Update CFGGuide to use outlines.fsm.parsing. Enable generate.cfg by @lapp0 in #1067
• Include hidden files in coverage CI upload by @brandonwillard in #1136
• Add documentation request issue template by @cpfiffer in #1138
• Set the tokenizer versions in test_create_fsm_index_tokenizer by @brandonwillard in #1139
• Change Outlines' logo by @rlouf in #1143
• Update logo size in documentation by @rlouf in #1144
• Improve sampler docs by @cpfiffer in #1141
• Update vllm.md by @americanthinker in #1137
• Correct pathways, update site color, front page fixes by @cpfiffer in #1146
• Change the color of the logo by @rlouf in #1155
• Remove Broken pyairports Package, Replace with airportsdata by @lapp0 in #1156
• Enable Tokenizers with Byte Tokens by @lapp0 in #1153
• Integrate OpenAI API Structured Generation by @lapp0 in #1142
• Remove link to Outlines twitter account by @rlouf in #1168
• Don't re-use logits processors in SequenceGeneratorAdapter, copy them by @lapp0 in #1160
• Fix benchmark workflow triggers by @brandonwillard in #1170
• Reuse jinja environment for a prompt by @jantrienes in #1162
• Use Faster FSM by @lapp0 in #1175
• Pin outlines-core version by @brandonwillard in #1187
• add missing comma in llamacpp docs by @cpfiffer in #1190
• Minor Changes on Top of Isamus EXL2 PR (Docs, Enable FSM/CFG, Fix Tokenizer) by @lapp0 in #1191
• Version the Documentation by @lapp0 in #1059

New Contributors

• @BIMM99 made their first contribution in #1004
• @robcaulk made their first contribution in #1008
• @willkurt made their first contribution in #1012
• @rshah713 made their first contribution in #1029
• @shinohara-rin made their first contribution in #1030
• @JulesGM made their first contribution in #1025
• @gtsiolis made their first contribution in #1034
• @talboren made their first contribution in #1036
• @JerryKwan made their first contribution in #1046
• @milo157 made their first contribution in #1047
• @Perdu made their first contribution in #1057
• @ispobock made their first contribution in #1068
• @tiran made their first contribution in #1094
• @cpfiffer made their first contribution in #1101
• @billmetangmo made their first contribution in #1120
• @americanthinker made their first contribution in #1137
• @jantrienes made their first contribution in #1162

Full Changelog: 0.0.46...0.1.0