Models.dev - An open-source AI model database from the OpenCode team.

What is Models.dev?

Models.dev is an open-source AI model database from the OpenCode team. Models.dev integrates model information from dozens of mainstream vendors, including OpenAI, Anthropic, and Google, in a standardized TOML format. It covers key specifications such as price, context window, tool calls, and knowledge cutoffs, and provides a public JSON API (models.dev/api.json) for developers to use directly. Models.dev addresses the pain point of fragmented AI model information, helping developers to perform one-stop queries and accurate cost estimations, and efficiently select models.

The main functions of Models.dev

• One-stop model information queryIt includes basic dialogue models and professional domain models from dozens of mainstream vendors such as OpenAI, Anthropic, Google, and Meta, eliminating the need to switch between multiple official websites to search.
• Accurate cost estimationIt provides detailed records of the input/output/inference/cache read/cache write costs per million tokens (USD), as well as audio input/output costs, helping projects accurately estimate budgets.
• Capability Specifications ComparisonThe standardized display model should support key indicators such as tool calls, reasoning/thought chains, structured output, temperature control, file attachments, knowledge deadlines, context windows, maximum number of input/output tokens, and supported modalities.
• Public JSON API:supply models.dev/api.json The interface allows developers to directly call complete model data in their projects to build model selection tools, cost calculators, or management panels.
• Obtain Provider Logo:pass models.dev/logos/{provider}.svg The interface directly retrieves the logos of various manufacturers.
• Community Collaboration MaintenanceData is stored in TOML format, categorized by provider, and community contributions are supported to ensure continuous database updates.

The technical principles of Models.dev

• Three-phase loop architectureThe input is encoded by Prelude (standard Transformer layer), then enters the Recurrent Block loop for T iterations, and is finally output by Coda; each loop prevents hidden state drift by injecting the original input e.
• Implicit mind chainEach loop is equivalent to one Chain-of-Thought reasoning step, but it runs silently in the continuous latent space and does not output intermediate tokens; it supports encoding multiple reasoning paths simultaneously.
• LTI stability constraintsTreating the loop as a linear time-invariant system, by parameterizing A as a negative diagonal matrix and discretizing it, we strictly guarantee that the spectral radius ρ(A) < 1, thus fundamentally solving the problems of residual explosion and training divergence.
• MoE+ Cyclic Synergistic EffectMoE provides breadth across domains, while the loop mechanism provides depth of reasoning; as the hidden state evolves in the loop, the router may select different subsets of experts, making each loop computation unique.
• Adaptive stoppingIt supports the ACT (Adaptive Computation Time) mechanism, where the model dynamically decides when to stop the loop, avoiding "overthinking".

How to use Models.dev

• Install:pip install open-mythosOptional[flash]Enable Flash Attention 2.
• Configuration:choosemlaorgqaAttention type, usingMythosConfigSet parameters such as dimension, number of heads, and number of loops.
• Create a model:OpenMythos(cfg)Initialize the network.
• Inference generation: callmodel.generate(max_new_tokens=8, n_loops=8),passn_loopsControl the depth of reasoning.
• train:usetraining/3b_fine_web_edu.pyScript, can run directly on a single card or through multiple cardstorchrunstart up.

Key information and usage requirements for Models.dev

• environmentPython + PyTorch; Flash Attention 2 requires CUDA and a compilation toolchain.
• Word segmenter:useopenai/gpt-oss-20bTokenizer.
• accuracyFor H100/A100, bfloat16 is recommended; for older GPUs, use float16 + GradScaler.
• Training configurationAdamW optimizer, linear warmup with cosine decay after 2000 steps, target approximately 30B tokens.
• Scale CoverageThere are predefined parameters for both the 1B experimental model and the 1T theoretical configuration.

The core advantages of Models.dev

• High efficiency of parametersA k-layer loop with L iterations is equivalent to a kL-layer fixed-depth network, with only the number of parameters of the k-layer type and memory usage that does not increase with inference depth.
• Reasoning is scalableIncreasing the number of iterations during testing can improve reasoning ability, following a predictable saturation decay pattern.
• Training stabilityThe LTI constraint mechanism completely solves the problems of instability and loss spikes in the training of recurrent models.
• System generalizationIt excels in out-of-distribution (OOD) combinatorial reasoning and achieves a leap in ability through a three-stage "insight" process.
• Depth extrapolationThe 5-step inference chain can be successfully expanded to 10 steps during testing, while the vanilla Transformer fails.

The project address for Models.dev

• Project official websitehttps://models.dev/
• GitHub repositoryhttps://github.com/anomalyco/models.dev

Comparison with similar products from Models.dev

Application scenarios of Models.dev

• AI Architecture Research: Verify the extended theory of looping Transformers, implicit inference chains, and test-time computation.
• Attention mechanism experimentCompare the KV caching efficiency and inference quality of MLA and GQA under a circular architecture.
• MoE research: Testing the impact of combining sparse expert routing with loop depth on multi-domain tasks.
• Model trainingBased on open-source code and scripts, train custom-scale models from scratch on datasets such as FineWeb-Edu.
• Stability StudyTo verify the practical effects of techniques such as LTI constraint, spectral radius control, and continuous depth batch processing.