Who is working on the ingredients of better AI models?
A selected map of established developers, specialist suppliers and open research projects. Inclusion reflects relevant technical work, not a ranking by revenue, valuation or market share. Follow each link to the organization’s published work.
Frontier language and multimodal models; capability at different operating budgets.
Multimodal foundation models, reasoning, and vision-language-action models for physical interaction.
Mixture-of-experts architectures, long context, and efficient model execution.
Claude model families, coding and professional capabilities, and model safety.
Generalist robot policies, action-model steerability, and cross-embodiment generalization.
Language and multimodal models, including specialized speech and translation model development.
Open model datasets, synthetic generation, curation
Open datasets, distribution, data collection hardware
Automated corpus curation and data mixtures
Expert training data, physical-AI data, evaluations
Expert data, work evaluations, post-training environments
Expert labeling and specialist-model RL
Human feedback and work evaluation
Expert training data, evaluations, training platform
Open model training recipes, low-precision pretraining and post-training tooling.
Training systems research and commercial customization infrastructure.
Tinker training API and published domain-specialization studies.
Major model developer whose self-serve training lifecycle changes affect specialization plans.
Important model-training research organization
Contributor affiliation on FlashAttention-4 and open training infrastructure
Cache-aware model routing, specialized models and hosted inference.
Hosted tool execution inside model inference; serving systems and model APIs.
Disaggregated serving, CPU frontends and predictable latency.
KV-cache precision, resident session capacity and quantization tradeoffs.
Routing, dedicated serving, rollout control, speculative decoding and program-aware scheduling.
Programmable execution and inference deployment.
FrontierCode evaluates code quality and mergeability; version 1.1 addresses solution lookup and grading false negatives. Also published July trustworthiness evaluations for open-source-derived models.
Legacy-Bench measures legacy-language tasks; September evaluation guidance connects evaluation to actual repository acceptance.
Controlled experiments quantify how resource allocations alter agentic coding scores.
GDP.xlsx evaluates professional spreadsheet reasoning, alongside a broader portfolio of expert-authored benchmarks.
