New AI Models

Phi-3.5-MoE-instruct

Phi-3.5-MoE-instruct is Microsoft's open weight (MIT license) mixture-of-experts (MoE) language model released in August 2024. It carries 16x3.8B parameters in total, activates 6.6B parameters per token by routing to 2 experts, and was trained with the GRIN (GRadient INformed MoE) method for expert specialization. The model supports 128K token context and focuses on multilingual quality, math, reasoning, and code generation. Microsoft reports that it outperforms models with similar or larger compute budgets such as Mistral-Nemo-12B, Llama-3.1-8B, and Gemma-2-9B, and matches Gemini-1.5-Flash in several categories. It is available via Serverless API in Azure AI Studio and GitHub, and suits applications that need strong reasoning and multilingual capability at a low active parameter cost.

Phi-3.5-vision-instruct

Phi-3.5-vision-instruct is the vision member of Microsoft's Phi-3.5 family, an open (MIT) multimodal model with 4.2B parameters that combines an image encoder, connector, and projector with the Phi-3 Mini language model, and supports a 128K token context. Released in August 2024, it covers general image understanding, OCR, chart and table understanding, multi-image comparison, and multi-frame video or slide summarization, making it a fit for memory and latency constrained environments. Microsoft reports single-image benchmark gains over Phi-3-vision, reaching MMMU 43.0, MMBench 81.9, and TextVQA 72.0, with BLINK overall at 57.0 and Video-MME at 50.8. It is trained mainly for English and is not optimized for multilingual use without further fine-tuning.

2.0

Ideogram 2.0, released on August 21, 2024, is a text-to-image model trained from scratch by Ideogram, with stated strengths in photorealism, graphic design, typography, and industry-leading text rendering. It offers distinct styles (Realistic, Design, 3D, Anime), any aspect ratio up to 3:1 and 1:3, and color palette control for consistent visual tones. The launch also introduced the Ideogram iOS app and a beta API. Per the vendor's human evaluations, it was rated ahead of Flux Pro and DALL·E 3 on image-text alignment, subjective preference, and text rendering accuracy. It suits posters, print-on-demand artwork, marketing visuals, and brand assets that need reliable in-image text.

V6.1

V6.1 is Midjourney's image generation model released on July 30, 2024, and it became the default model for all users with this release. Compared with V6, it renders more coherent arms, legs, hands and bodies, improves overall image quality with fewer pixel artifacts and richer textures, and captures small features such as eyes, small faces and distant hands more precisely. It also introduces new 2x upscalers with better texture quality, more accurate text drawn from quoted words in prompts, and a new personalization model with personalization code versioning. Standard image jobs run roughly 25% faster, and an optional high-texture quality mode takes 25% longer to add detail at some cost to coherence. It suits artistic image creation, concept design and illustration workflows driven by prompt aesthetics; regional editing still falls back to the V6.0 model.

Mistral Large 2

Mistral Large 2 is the flagship open-weight LLM Mistral AI released in July 2024, a dense 123B parameter model built for single-node inference with a 128k token context window. It supports dozens of languages including English, French, German, Spanish, Chinese, Japanese, Korean, Arabic and Hindi, and codes in 80+ programming languages such as Python, Java, C++, JavaScript and Bash. According to Mistral, the pretrained version reaches 84.0% on MMLU, and the model performs on par with GPT-4o, Claude 3 Opus and Llama 3 405B on code generation and reasoning benchmarks. Training also emphasized lower hallucination, concise outputs, and parallel and sequential function calling. Weights are available on Hugging Face under the Mistral Research License (API name mistral-large-2407), suiting long-context, multilingual and code-heavy applications; commercial self-deployment requires a Mistral Commercial License.

Llama 3.1 405B Instruct

Llama 3.1 405B Instruct is Meta's flagship open instruction-tuned LLM, released on July 23, 2024 as the largest model in the Llama 3.1 family (8B, 70B, 405B). Built on an optimized transformer architecture with Grouped-Query Attention, it was pretrained on over 15 trillion tokens of publicly available data, supports a 128K context window, and officially covers 8 languages including English, German, French, Italian, Portuguese, Spanish, Hindi, and Thai. Official benchmark results for the 405B Instruct model include MMLU 87.3, HumanEval 89.0, GSM-8K 96.8, MATH 73.8, multilingual MGSM 91.6, and strong tool calling scores such as API-Bank 92.0 and BFCL 88.5. Meta states the models outperform many available open source and closed chat models. Licensed for commercial and research use, including synthetic data generation and distillation, it fits assistant-like chat applications that need long-context reasoning, coding, math, and multilingual generation with openly available weights.

Llama 3.1 70B Instruct

Llama 3.1 70B Instruct is the 70B-parameter instruction-tuned model in Meta's open Llama 3.1 family, released on July 23, 2024 under the Llama 3.1 Community License. It is an auto-regressive transformer aligned with SFT and RLHF, uses Grouped-Query Attention (GQA) for more scalable inference, and supports a 128K context window across 8 languages (English, German, French, Italian, Portuguese, Hindi, Spanish, Thai). Beyond assistant-style chat and multilingual dialogue, it covers general knowledge, reasoning, coding, math, and tool calling; official benchmark scores include MMLU 86.0 (CoT), MMLU-Pro 66.4, IFEval 87.5, HumanEval 80.5, GSM-8K 95.1, MATH 68.0, and BFCL 84.8 for tool use. It suits multilingual chat assistants, agent workflows, and cost-effective self-hosted deployments where a single model must balance quality and serving cost.

Llama 3.1 8B Instruct

Llama 3.1 8B Instruct is Meta's instruction-tuned, text-only model and the smallest member of the Llama 3.1 open family, released on July 23, 2024 under the Llama 3.1 Community License. It has 8B parameters, a 128K token context window, and official support for 8 languages (English, German, French, Italian, Portuguese, Hindi, Spanish, Thai). Pretrained on roughly 15 trillion tokens and aligned with SFT and RLHF using over 25M synthetic examples, it adds tool calling and multilingual, code, math, and long-context abilities. The official Model Card reports MMLU 69.4, HumanEval 72.6, GSM-8K (CoT) 84.5, IFEval 80.4, and BFCL 76.1, making it a practical choice for cost-sensitive chat assistants, coding helpers, and agent workflows on a single GPU.

Mistral NeMo Instruct

Mistral NeMo Instruct is the instruction-tuned version of Mistral NeMo, a 12B open-weight LLM released on 2024-07-18 by Mistral AI in collaboration with NVIDIA under the Apache 2.0 license. It offers a 128k token context window, function calling, and strong performance across English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic and Hindi. Its Tekken tokenizer, based on Tiktoken and trained on over 100 languages, compresses source code and several major languages about 30% more efficiently than the earlier SentencePiece tokenizer, and outperforms the Llama 3 tokenizer on roughly 85% of languages. Trained with quantization awareness, it runs FP8 inference without performance loss and serves as a drop-in replacement for Mistral 7B. The official model card reports MMLU (5-shot) at 68.0%, HellaSwag (0-shot) at 83.5% and Winogrande (0-shot) at 76.8%. It suits multilingual chat, coding assistance, long-document workflows and agents that rely on function calling.

GPT-4o mini

GPT-4o mini is OpenAI's cost-efficient small multimodal model, released on July 18, 2024 as a more affordable replacement for GPT-3.5 Turbo. It accepts text and image input, offers a 128K token context window with up to 16K output tokens per request, and shares GPT-4o's improved tokenizer so non-English text costs less to process. On official benchmarks it scores 82.0% on MMLU, 87.0% on MGSM, 87.2% on HumanEval and 59.4% on MMMU, and outperforms GPT-4 Turbo 01-25 on LMSYS chat preferences, with video and audio support to follow. At $0.15 per 1M input tokens and $0.60 per 1M output tokens, it suits high-volume workloads such as chained or parallel model calls, long-context tasks, function calling, and real-time customer support chatbots.

Eleven Flash v2.5

Eleven Flash v2.5 (model id eleven_flash_v2_5) is ElevenLabs' fastest text-to-speech (TTS) model, designed for real-time applications. Official documentation lists ultra-low latency of about 75ms, excluding application and network latency, across 32 languages, adding Hungarian, Norwegian and Vietnamese on top of the Multilingual v2 language set. Each request accepts up to 40,000 characters, and API generations come at a 50% lower price per character. The model balances speed and quality with natural-sounding, consistent voices, and ElevenLabs recommends it for real-time voice agents and chatbots, interactive apps such as games, and large-scale bulk speech generation. Docs also note that Flash models have lower average latency than their Turbo counterparts, with Flash recommended over Turbo in all use cases.

Eleven Turbo v2.5

Eleven Turbo v2.5 is ElevenLabs' low-latency text-to-speech (TTS) model released in July 2024 as an upgrade to the Turbo line. According to the official release, it supports 32 languages, runs 3x faster on Hindi, French, Spanish, Mandarin and 27 other languages, and is 25% faster in English, while adding Vietnamese, Hungarian and Norwegian for the first time. Aimed at conversational AI and other real-time voice applications, it targets use cases where low latency and broad language coverage matter, and ElevenLabs states it brings high-quality low-latency voice to nearly 80% of the world.

kling-v1-5

kling-v1-5 is an early-generation Kling video generation model on the KlingAI open platform. It provides image-to-video generation through the /v1/videos/image2video API: standard mode (std) outputs 720P video, while the professional mode (pro) outputs 1080P for higher quality. Official release notes show it was opened for image-to-video only, with text-to-video not supported, and it later gained start-and-end frame control, motion brush, camera control, lip sync for generated clips, and video extension that continues a clip by about 4 to 5 seconds. Developers can steer results with prompt and negative_prompt parameters. It now serves as a legacy option below Kling 1.6 and the newer 2.x and 3.x models, and the official docs mark it for retirement in favor of mainline models.

CriticGPT

CriticGPT is an OpenAI research model built on GPT-4, announced June 27, 2024, and trained to write critiques that surface mistakes in ChatGPT's code output. It was trained with RLHF on code where human trainers deliberately inserted bugs and then wrote feedback for them. In official experiments, human reviewers assisted by CriticGPT outperformed unassisted reviewers in 60% of comparisons, and CriticGPT critiques were preferred over ChatGPT self-critiques in 63% of cases on naturally occurring bugs, with fewer nitpicks and fewer hallucinated problems. OpenAI plans to integrate CriticGPT-like models into its RLHF labeling pipeline, so it suits human evaluation of model outputs and code error detection, though the team notes limits on long answers, dispersed errors, and residual hallucinations.

Claude 3.5 Sonnet

Claude 3.5 Sonnet is the first release in Anthropic's Claude 3.5 family, launched on June 21, 2024. It outperforms the previous flagship Claude 3 Opus across evaluations while running at twice the speed of Opus, and keeps mid tier pricing at $3 per million input tokens and $15 per million output tokens with a 200K token context window. The model set new industry records on graduate-level reasoning (GPQA), undergraduate knowledge (MMLU), and coding (HumanEval), and in an internal agentic coding evaluation it solved 64% of problems versus 38% for Claude 3 Opus. Anthropic also describes it as its strongest vision model to date. The same launch introduced Artifacts, a preview feature on Claude.ai where generated code snippets, documents, and website designs open in a dedicated window beside the chat for real-time editing. It suits complex coding, data analysis, visual understanding, and general enterprise work, and is available on Claude.ai and the Claude iOS app as well as through the Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI. Claude 3.5 Haiku and Claude 3.5 Opus were announced to follow later that year.

kling-v1

kling-v1 is the first-generation video generation model of Kuaishou's Kling platform (Kling 1.0), launched in June 2024. Per the official Kling 1.0 API docs, kling-v1 is the default model name and supports text-to-video and image-to-video generation, plus video extension. Text-to-video takes a prompt of up to 2500 characters with negative prompt, cfg_scale, and camera movement controls such as pan, tilt, roll, and zoom; image-to-video animates a reference image and can combine a start frame with an optional end frame. Output runs in std mode (720p) or pro mode (1080p) at 16:9, 9:16, or 1:1, and Kuaishou positioned the launch version around videos of up to about two minutes at 1080p. It fits short-form content, marketing clips, and adding motion to still images; the official docs note the model will be retired from the API on 2026-09-15.

Codestral-22B

Codestral-22B is the first code model from Mistral AI, an open-weight 22B generative model designed for code generation tasks. Trained on 80+ programming languages including Python, Java, C, C++, JavaScript and Bash, it serves instruct and fill-in-the-middle (FIM) requests through a shared API endpoint, completing coding functions, writing tests and filling in partial code. With a 32k context window, Mistral reports that it outperforms other models in RepoBench, a long-range repository-level completion eval, and the official release also compares it on HumanEval, MBPP, CruxEval, Spider and FIM benchmarks against code models such as DeepSeek Coder 33B. Licensed under the Mistral AI Non-Production License (MNPL) for research and testing, with commercial licensing available on demand, it targets IDE code completion, coding assistants and code-generation applications, with integrations in LlamaIndex, LangChain, Continue.dev and Tabnine.

Golden Gate Claude

Golden Gate Claude is an interpretability research demo from Anthropic, not a production model. Built on Claude 3 Sonnet, it locates the Golden Gate Bridge feature identified by sparse autoencoder research and clamps it to about 10x its maximum activation, a precise change to the model's internal activations rather than prompting or fine-tuning. As a result the model brings up the bridge in nearly every reply, from suggesting a drive across it to writing a love story about it. It was briefly available on claude.ai for a 24 hour period in May 2024 and serves to demonstrate how feature level interpretability could one day help steer model behavior and improve safety.

GPT-4o

GPT-4o is OpenAI's flagship natively multimodal model released on May 13, 2024, where o stands for omni. A single neural network trained end-to-end accepts any combination of text, audio, image, and video as input and generates text, audio, or image outputs, responding to audio in as little as 232 milliseconds with an average of 320 milliseconds, close to human conversational pace. Official evaluations put it at GPT-4 Turbo level on text, reasoning, and coding while setting new highs on multilingual, audio, and vision understanding, and its new tokenizer needs far fewer tokens for many non-English languages. The API runs twice as fast as GPT-4 Turbo at half the cost with 5x higher rate limits, and the model is open to free ChatGPT users. It suits real-time voice assistants, image understanding, and multilingual chat products.

DeepSeek-V2.5

DeepSeek-V2.5 is an upgraded Mixture-of-Experts (MoE) model from DeepSeek that merges DeepSeek V2 Chat and DeepSeek Coder V2 into a single model built on the V2 architecture (236B total parameters, 21B activated per token, 128K context), using Multi-head Latent Attention (MLA) and DeepSeekMoE for efficient inference. Per the official release notes, it clearly surpasses its predecessors in both general and coding abilities, with stronger alignment in writing and instruction following: ArenaHard win rate rises from 68.3% to 76.3%, MT-Bench from 8.84 to 9.02, HumanEval reaches 89%, and LiveCodeBench (January to September) reaches 41%. API users can call it through either deepseek-chat or deepseek-coder for backward compatibility. It suits general chat, coding assistance, and reasoning-heavy workloads.