210 results
Foundation models you can call or deploy
by Meta·Apr 2024
Meta's 8B dense open-weight LLM from the Llama 3 family with an 8K token context, released April 2024 in base and Instruct variants, Llama 3 Community License.
by Anthropic·May 2025
Anthropic's Claude Sonnet 4, initial Claude 4 mid-tier model for coding and reasoning; launched alongside Opus 4, retired June 15, 2026.
by Mistral AI·May 2025
Mistral's 24B open-weight agentic coding model, Apache 2.0, tuned for tool use in software-engineering tasks; built with All Hands AI, tops open-model SWE-Bench.
by Alibaba·Jan 2025
Vision-language Qwen2.5 family (3B/7B/72B) with dynamic-resolution ViT and 32K context (YaRN extendable), qwen license; strong OCR and video grounding.
by Meta·Sep 2024
Meta's 3B dense open-weight text model from Llama 3.2 with a 128K token context, optimized for edge and mobile inference, Llama 3.2 Community License.
by Meta·Jul 2024
Meta's flagship 405B dense open-weight LLM from Llama 3.1 with a 128K token context, released July 2024 as the first frontier-scale open-weights model, Llama 3.1 license.
by Z.ai·Mar 2026
Faster turbo variant of Z.ai's GLM-5, optimized for agentic coding workflows; MoE with sparse attention. Announced/rumored tier; release details limited.
by Meta·Dec 2024
Meta's Llama 3.3 70B Instruct-only dense open-weight LLM with 128K context, delivering 405B-class quality at 70B cost, under the Llama 3.3 Community License.
by Mistral AI·May 2026
Mistral's frontier mid-size multimodal model optimized for agentic and coding workflows; default model in Vibe and Le Chat, released as open weights under a modified MIT license.
by Anthropic·Feb 2025
Anthropic's hybrid-reasoning Claude model with toggleable extended thinking; reached state-of-the-art scores on SWE-bench Verified at launch.
by Meta·Apr 2025
Meta's Llama 4 Maverick MoE model with 17B active / 128 experts (400B total), 1M token context, natively multimodal, under the Llama 4 Community License.
by Moonshot AI·May 2026
1T/32B-active MoE with 256K context, MoonViT vision, Modified MIT license; coding-focused successor to K2.5 with 300-agent swarm and long-horizon coding.
by Z.ai·Feb 2026
744B/40B-active MoE (glm_moe_dsa) with DeepSeek Sparse Attention, MIT license; trained on 28.5T tokens for long-horizon agentic engineering tasks.
by Anthropic·Jun 2026
Anthropic's Claude Mythos 5 frontier model for defensive cybersecurity, offered in limited availability to Project Glasswing partners; 1M-token context.
1B multimodal OCR model, MIT license; complex document understanding with layout grounding and multilingual text extraction for real-world docs.
Meta's 70B dense open-weight LLM from the Llama 3 family with an 8K token context, released April 2024 in base and Instruct variants, Llama 3 Community License.
Meta's Llama 3.2 open-weight family: 1B and 3B text models with 128K context plus 11B and 90B vision multimodal variants, under the Llama 3.2 Community License.
by xAI·Jul 2025
xAI's flagship model succeeding Grok 3; adds native tool use and real-time search, trained with massively scaled reinforcement learning on the 200k-GPU Colossus cluster.
by xAI·Feb 2025
Smaller, cost-efficient sibling of Grok 3 launched alongside it; reasoning-focused variant priced near GPT-4o-mini and noted for strong coding performance.
by DeepSeek·Dec 2025
685B MoE variant of V3.2 with DSA, MIT license; reasoning-only tier with gold-medal IMO 2025/IOI 2025 performance and no tool-calling support.
by Mistral AI·Sep 2024
Mistral's first multimodal vision-language model: 12B decoder plus 400M vision encoder, 128k context, Apache 2.0; ingests images of arbitrary sizes and counts.
by Microsoft·Apr 2024
Microsoft's 3.8B dense Phi-3 SLM with 4K or 128K context variants (MIT license), tuned to punch above its size on reasoning and instruction-following benchmarks.
by OpenAI·Feb 2019
OpenAI's 2019 transformer language model (up to 1.5B parameters), known for coherent zero-shot text generation; initially released in staged increments over misuse concerns.
by Mistral AI·Sep 2023
Mistral's 7.3B dense open-weight LLM under Apache 2.0; uses grouped-query and sliding-window attention, a widely fine-tuned baseline for local inference.
