Large Language Models (LLMs)

Versatile multilingual embedding model supporting dense, sparse, and multi-vector retrieval

Provider: BAAI · Context window: 8K · Parameters: 568M

GPT-5 is OpenAI's August 2025 reasoning model for coding, reasoning, and agentic tasks. The API model supports text input and output, image input, configurable reasoning effort, and an immutable gpt-5-2025-08-07 snapshot; OpenAI now classifies it as a previous-generation model.

Provider: OpenAI · Context window: 400K

GPT-5 Mini is OpenAI's August 2025 cost-efficient reasoning model for well-defined, latency-sensitive, and high-volume workloads. The API model supports text input and output, image input, configurable reasoning effort, and an immutable gpt-5-mini-2025-08-07 snapshot; OpenAI now classifies it as a previous-generation model.

Provider: OpenAI · Context window: 400K

Claude Sonnet 4.5 is Anthropic's balanced frontier model, designed for complex reasoning, coding, and agentic tasks while maintaining production-grade speed. It builds on the Claude 4 family's emphasis on honesty and reduced hallucination rates.

Provider: Anthropic · Context window: 200K

Claude Opus 4 is Anthropic's May 2025 hybrid-reasoning model for complex coding, reasoning, multimodal analysis, and long-running agent workflows. The dated claude-opus-4-20250514 snapshot was retired on Anthropic-operated platforms on June 15, 2026 and is retained as a historical evaluation and migration reference.

Provider: Anthropic · Context window: 200K

Gemini 3 Pro is Google DeepMind's flagship multimodal model with a 2M-token context window spanning text, image, audio, and video inputs. It is built for long-context reasoning, document understanding, and multimodal agentic workflows.

Provider: Google · Context window: 2M

Llama 4 Scout is Meta's efficiency-focused open-weights model with an exceptionally large 10M-token context window and vision support. It is optimized for long-document retrieval and low-cost inference at scale.

Provider: Meta · Context window: 10M

DeepSeek R1 is DeepSeek's open-source reasoning model, trained with reinforcement learning to produce chain-of-thought reasoning before answering. Released in January 2025 under the MIT license, it matched or exceeded closed reasoning models on several math and coding benchmarks.

Provider: DeepSeek · Context window: 128K

Mistral Large 2 is Mistral AI's 2024 flagship commercial model with a 128,000-token context window, built for instruction following, multilingual tasks and function calling in agentic workflows. Both releases are now retired on Mistral's own API: mistral-large-2407 on 30 March 2025, replaced by Mistral Large 3, and mistral-large-2411 on 31 May 2026, replaced by Mistral Medium 3.5.

Provider: Mistral AI · Context window: 128K

Command R+ is Cohere's open-weights enterprise model optimized for retrieval-augmented generation, tool use, and multilingual instruction-following across 10 languages. Released under CC-BY-NC, it targets production RAG pipelines and enterprise search workflows.

Provider: Cohere · Context window: 128K

o3 is OpenAI's April 2025 reasoning model, applying deliberate chain-of-thought computation at inference time to mathematics, science and competitive programming tasks, and the release that brought OpenAI's scalable test-time compute approach to its production API. OpenAI has since deprecated it, scheduling the o3-2025-04-16 snapshot to shut down on 11 December 2026 with gpt-5.6-sol as the replacement.

Provider: OpenAI · Context window: 200K

o4 Mini is OpenAI's efficient reasoning model, adding vision support to the mini reasoning tier for agentic tasks needing both image understanding and step-by-step logic at low cost per task. It is now deprecated: OpenAI gives the o4-mini-2025-04-16 snapshot and the o4-mini alias a shutdown date of 23 October 2026, with gpt-5.6-terra as the substitute.

Provider: OpenAI · Context window: 200K

Text Embedding 3 Large is OpenAI's larger third-generation embedding model, returning 3,072-dimension vectors from inputs of up to 8,192 tokens. Its dimensions parameter shortens the vector at creation time without, as OpenAI puts it, the embedding losing its concept-representing properties, which lets a retrieval system trade index size against accuracy without changing models. Shortening at creation is preferred to truncating afterwards, which requires re-normalising the vector.

Provider: OpenAI · Context window: 8K

Claude Sonnet 4 is Anthropic's May 2025 hybrid-reasoning model for coding, visual reasoning, tool use, and agent workflows. The dated claude-sonnet-4-20250514 snapshot was retired on Anthropic-operated platforms on June 15, 2026 and is retained as a historical evaluation and migration reference.

Provider: Anthropic · Context window: 200K

Claude 3.5 Sonnet is Anthropic's mid-2024 balanced model, the release that introduced computer use and anchored Anthropic's agentic-coding positioning for its generation across a 200,000-token context window. It is now retired: Anthropic deprecated both the claude-3-5-sonnet-20240620 and claude-3-5-sonnet-20241022 snapshots on 13 August 2025 and retired them on 28 October 2025 on Anthropic-operated platforms, naming Claude Sonnet 5.5 as the replacement.

Provider: Anthropic · Context window: 200K

Claude 3 Haiku is the speed-and-cost tier of Anthropic's original Claude 3 family, built for high-volume work such as customer support, document triage and many small tasks in parallel, with image input across a 200,000-token context window. Anthropic measured it at 21,000 tokens per second for prompts under 32,000 tokens. The claude-3-haiku-20240307 snapshot was deprecated on 19 February 2026 and retired on Anthropic-operated platforms on 20 April 2026, with Claude Haiku 4.5 as the replacement.

Provider: Anthropic · Context window: 200K

Gemini 2.5 Pro is Google DeepMind's experimental thinking model with a 2M-token multimodal context window, released in March 2025. It topped the LMSys Chatbot Arena leaderboard at launch and introduced adaptive reasoning budgets that balance response quality against latency.

Provider: Google · Context window: 2M

Gemini 2.5 Flash is Google DeepMind's price-performance model for low-latency, high-volume work that still needs reasoning, with thinking supported. It accepts text, image, audio and video input across a 1,048,576-token context window, returns up to 65,536 tokens, and carries a January 2025 knowledge cutoff. Google now limits access to existing users while keeping the model served.

Provider: Google · Context window: 1M

Gemma 3 27B is the largest model in Google's third-generation open Gemma family, accepting text and images across a 128,000-token input context, with output drawn from the same budget. Images are normalised to 896x896 and encoded as 256 tokens each. Google cites support for more than 140 languages and an August 2024 knowledge cutoff. The weights are published under Google's Gemma Terms of Use rather than a standard open-source licence.

Provider: Google · Context window: 128K

Llama 3.1 405B Instruct is the largest model in Meta's Llama 3.1 family and the one that made open weights available at frontier scale. It handles a 128,000-token context, officially supports eight languages, and was trained on roughly 15 trillion tokens with a December 2023 knowledge cutoff. It ships under the Llama 3.1 Community License, which carries conditions rather than being a standard open-source licence.

Provider: Meta · Context window: 128K

Llama 3.1 70B Instruct is the mid-size model in Meta's Llama 3.1 family, sharing the family's 128,000-token context, eight officially supported languages, roughly 15 trillion training tokens and December 2023 knowledge cutoff. It is the size most often chosen when 405B is impractical to host and 8B is not capable enough. It ships under the Llama 3.1 Community License rather than a standard open-source licence.

Provider: Meta · Context window: 128K

Llama 3.1 8B Instruct is the smallest model in Meta's Llama 3.1 family, carrying the same 128,000-token context, eight officially supported languages and December 2023 knowledge cutoff as its larger siblings. Its size makes it the family member most often fine-tuned or run on a single GPU. It ships under the Llama 3.1 Community License, which carries conditions rather than being a standard open-source licence.

Provider: Meta · Context window: 128K

Mistral Small 3.1 is Mistral AI's 24-billion-parameter multimodal model, accepting text and image input across a context window of up to 128,000 tokens and published under Apache 2.0. Mistral reports inference speeds of 150 tokens per second and compares it with Gemma 3 and GPT-4o Mini in the same weight class. Its size and licence make it one of the larger models that remains practical to self-host on a single machine.

Provider: Mistral AI · Context window: 128K

DeepSeek V3 is DeepSeek's December 2024 open-weights Mixture-of-Experts model, activating 37 billion of 671 billion total parameters per token across a 128,000-token context window. It ships under the MIT licence with published weights and a technical report, and DeepSeek reports its own evaluation table for the text chat model under an 8,000-token output limit.

Provider: DeepSeek · Context window: 128K

QwQ 32B is the reasoning model of Alibaba's Qwen series, a 32.5-billion-parameter causal language model (31.0 billion excluding embeddings) with 64 layers and grouped-query attention using 40 query and 8 key-value heads. It is built to work through a problem before answering rather than responding directly. The full context length is 131,072 tokens, though Qwen notes that prompts beyond 8,192 tokens require YaRN to be enabled. Released under Apache 2.0.

Provider: Alibaba · Context window: 131K

Nemotron-4 340B Instruct is NVIDIA's large decoder-only transformer, released under the NVIDIA Open Model License and built primarily to generate synthetic training data for other models rather than to serve end users. NVIDIA reports that over 98 percent of the data used in its own alignment was synthesised, from roughly 20,000 human-annotated examples. Its 4,096-token context is short by current standards and is the main constraint on reusing it as a general assistant.

Provider: NVIDIA · Context window: 4K

Microsoft's 14B dense decoder-only open-weight text model, trained on 9.8T tokens with a synthetic-data-centered curriculum and available through the microsoft/phi-4 artifact and Microsoft Foundry.

Provider: Microsoft · Context window: 16K

GPT-4o is OpenAI's May 2024 multimodal flagship, accepting text and image input across a 128,000-token context window with a 16,384-token maximum output and an October 2023 knowledge cutoff. OpenAI describes it as its most capable model outside the o-series reasoning line, and serves it through dated snapshots with gpt-4o-2024-08-06 as the default. The family is now deprecated on two schedules: gpt-4o-2024-08-06 and gpt-4o-2024-11-20 shut down on 14 January 2027 with gpt-6-luna as the replacement, and the original gpt-4o-2024-05-13 snapshot on 23 October 2026 with gpt-5.6-sol.

Provider: OpenAI · Context window: 128K

GPT-4o Mini is OpenAI's small multimodal model, accepting text and image input across a 128,000-token context window with a 16,384-token maximum output and an October 2023 knowledge cutoff, priced at $0.15 per million input tokens. OpenAI has since deprecated it, scheduling the gpt-4o-mini-2024-07-18 snapshot to shut down on 14 January 2027 with gpt-6-luna as the replacement.

Provider: OpenAI · Context window: 128K

GPT-4.1 is OpenAI's April 2025 API-only model, built for instruction following and tool calling without a reasoning step. It accepts text and image input across a 1,047,576-token context window, returns up to 32,768 tokens, and carries a June 2024 knowledge cutoff — roughly eight times GPT-4o's context at lower per-token pricing. OpenAI has since deprecated it, scheduling the gpt-4.1-2025-04-14 snapshot to shut down on 14 January 2027 with gpt-6-luna as the replacement.

Provider: OpenAI · Context window: 1M

GPT-4.1 Mini is OpenAI's April 2025 smaller, faster GPT-4.1 variant, carrying the same 1,047,576-token context window, 32,768-token maximum output and June 2024 knowledge cutoff as the full model at roughly a fifth of its input price. It accepts text and image input and, like GPT-4.1, answers without a reasoning step. OpenAI has since deprecated it, scheduling the gpt-4.1-mini-2025-04-14 snapshot to shut down on 14 January 2027 with gpt-6-luna as the replacement.

Provider: OpenAI · Context window: 1M

Text Embedding 3 Small is OpenAI's smaller third-generation embedding model, returning 1,536-dimension vectors from inputs of up to 8,192 tokens at a fifth the price of text-embedding-ada-002. Like the larger third-generation model it accepts a dimensions parameter, which shortens the vector at creation time without the embedding losing its concept-representing properties, so index size can be traded against accuracy without changing models.

Provider: OpenAI · Context window: 8K

Claude 3.5 Haiku is Anthropic's fast, low-cost model from late 2024, built for high-volume work such as classification, extraction and routing across a 200,000-token context window. The claude-3-5-haiku-20241022 snapshot was deprecated on 19 December 2025 and retired on Anthropic-operated platforms on 19 February 2026, with Claude Haiku 4.5 named as its replacement.

Provider: Anthropic · Context window: 200K

Claude 3 Opus is Anthropic's February 2024 frontier model and the most capable of the original Claude 3 family, accepting text and image input across a 200,000-token context window. The claude-3-opus-20240229 snapshot was deprecated on 30 June 2025 and retired on Anthropic-operated platforms on 5 January 2026, with Claude Opus 4.8 named as its replacement.

Provider: Anthropic · Context window: 200K

Claude 3.7 Sonnet is Anthropic's February 2025 hybrid reasoning model, combining immediate answers with an optional extended thinking mode in a single model rather than two. Thinking can be budgeted by token count up to a 128,000-token output limit, across a 200,000-token context window. The claude-3-7-sonnet-20250219 snapshot was deprecated on 28 October 2025 and retired on Anthropic-operated platforms on 19 February 2026.

Provider: Anthropic · Context window: 200K

Gemini 2.0 Flash is Google's multimodal workhorse, accepting text, image, audio and video input across a 1,048,576-token context window with an 8,192-token output limit and an August 2024 knowledge cutoff. Google's deprecations page gives it a shutdown date of 1 June 2026 and names gemini-3.6-flash as the replacement, and the Gemini API models page now lists it among previous models as shut down; it is retained here as a historical reference.

Provider: Google · Context window: 1M

Gemini 1.5 Pro by Google.

Provider: Google · Context window: 2M

Gemini 1.5 Flash by Google.

Provider: Google · Context window: 1M

Mistral NeMo is a 12-billion-parameter model Mistral AI built with NVIDIA, handling up to a 128,000-token context and released under Apache 2.0 for both the base and instruction-tuned checkpoints. It introduced Tekken, a Tiktoken-based tokenizer trained on more than 100 languages that replaced the SentencePiece tokenizer of earlier Mistral models, with particularly strong coverage across eleven of them.

Provider: Mistral AI · Context window: 128K

Qwen 2.5 72B Instruct is the largest dense model in Alibaba's Qwen2.5 generation, 72.7 billion parameters (70.0 billion excluding embeddings) across 80 layers with grouped-query attention using 64 query and 8 key-value heads. It handles a full 131,072-token context and generates up to 8,192 tokens, with multilingual support Qwen puts at more than 29 languages. The weights are published under the Qwen licence rather than a standard open-source licence.

Provider: Alibaba · Context window: 131K

Qwen 2.5 Coder 32B Instruct is Alibaba's code-specialised model, 32.5 billion parameters (31.0 billion excluding embeddings) trained on 5.5 trillion tokens of source code, text-code grounding and synthetic data, and released under Apache 2.0. Qwen builds it for code generation, code reasoning and code repair. The published config sets a 32,768-token context by default; reaching the full 131,072 tokens requires YaRN scaling, which Qwen notes can degrade quality on shorter inputs.

Provider: Alibaba · Context window: 131K

Command R is Cohere's retrieval-focused model, built for conversational interaction and long-context tasks across a 128,000-token window, with capabilities in ten key languages and pre-training data covering thirteen more. Cohere positions it for retrieval-augmented generation and tool use, where it returns a JSON-formatted list of actions to run. Cohere deprecated the command-r alias together with the command-r-03-2024 snapshot on 15 September 2025, recommending command-r-08-2024, command-r-plus-08-2024 or command-a-03-2025.

Provider: Cohere · Context window: 128K

Text Embedding Ada 002 is OpenAI's second-generation embedding model, returning 1,536-dimension vectors. OpenAI now classes it as an older embedding model and prices it at five times text-embedding-3-small, which returns vectors of the same width. Its practical role is maintaining existing indexes: embeddings from different models are not comparable, so moving to a third-generation model means re-embedding the corpus rather than mixing vectors.

Provider: OpenAI · Context window: 8K

Mixtral 8x7B is Mistral AI's sparse mixture-of-experts model: 46.7 billion total parameters, of which a router activates only 12.9 billion per token by selecting two of eight expert groups at each layer. That gives it the inference cost of a far smaller dense model at the capacity of a large one. It handles a 32,000-token context and ships with open weights under Apache 2.0.

Provider: Mistral AI · Context window: 33K

Qwen2.5-VL 72B Instruct is the largest vision-language model in Alibaba's Qwen2.5-VL line, accepting images and video alongside text. Qwen highlights comprehension of videos over an hour long, structured extraction from documents such as invoices, forms and tables, visual grounding returned as bounding boxes or points, and agentic computer and phone use. The published config sets a 32,768-token context, and Qwen advises against YaRN extension here because it degrades temporal and spatial localisation.

Provider: Alibaba · Context window: 128K

InternVL2.5 78B is OpenGVLab's multimodal model in the ViT-MLP-LLM configuration, pairing the InternViT-6B-448px-V2.5 vision encoder with Qwen2.5-72B-Instruct as its language model for roughly 78 billion parameters in total. It keeps InternVL 2.0's architecture and advances instead through training and test-time strategy and a larger, filtered data mixture. The project is released under the MIT licence, while the Qwen2.5-72B-Instruct component it embeds carries the Qwen licence.

Provider: OpenGVLab · Context window: 33K

Pixtral 12B is Mistral AI's first multimodal model, pairing a 12-billion-parameter decoder based on Mistral NeMo with a 400-million-parameter vision encoder trained from scratch, under Apache 2.0. Its 128,000-token context can hold any number of images interleaved with text, and Mistral designed it as a drop-in replacement for Mistral NeMo 12B that adds image and document understanding without weakening text. Mistral now lists it as deprecated and no longer maintained.

Provider: Mistral AI · Context window: 128K