New AI Models

Qwen3.5-Plus-2026-04-20

Qwen3.5-Plus-2026-04-20 is the April 20, 2026 snapshot of qwen3.5-plus, the Plus-tier model in the Qwen3.5 native vision-language series, suited to production deployments that need stable model behavior. Compared with the February 15 snapshot, it offers stronger agentic coding and noticeably faster inference, while keeping knowledge, reasoning, and long-context capability high with a 1M context window. Recommended for coding agents, production workflows, and high-throughput scenarios.

bfl.ai

bfl.ai is the official platform of Black Forest Labs, the company behind the FLUX family of image generation and editing models. It serves FLUX.2 [pro] and FLUX.2 [flex] through a production API, alongside the 32B open-weight FLUX.2 [dev] checkpoint for self-hosting. The models cover text-to-image and multi-reference editing with reliable text rendering, accurate prompt following, and output up to 4MP. Access is available via the BFL API, open weights on Hugging Face and GitHub, and a browser-based Playground.

hy3-preview

Hy3-preview is a preview large language model from Tencent's Hunyuan team and the first trained on its rebuilt infrastructure. It uses a Mixture-of-Experts (MoE) architecture with 295B total parameters, 21B active parameters, and a 256K context window. Its biggest gains come in coding and agent tasks, with competitive scores on SWE-bench Verified, Terminal-Bench 2.0, and BrowseComp. Model weights are open-sourced under the Tencent Hy Community License on Hugging Face and ModelScope.

OpenAI Privacy Filter

OpenAI Privacy Filter is an open-weight model that detects and redacts personally identifiable information (PII) in unstructured text. Built as a bidirectional token classifier with span decoding, it offers a 128,000-token context window, labels eight PII categories in a single forward pass, and runs locally so sensitive data never leaves the machine. It reaches an F1 of 97.43% on the corrected PII-Masking-300k benchmark and is available under Apache 2.0 on Hugging Face and GitHub for privacy pipelines such as training data preparation, logging, and indexing.

Qwen3.6-27B

Qwen3.6-27B is a 27B native vision-language dense model in Alibaba's Qwen3.6 series, supporting text generation, deep thinking, and vision understanding. Compared with 3.5-27B, it delivers stronger agentic coding plus further gains in STEM and reasoning. On the vision side, spatial intelligence and object localization and detection are notably improved, and video understanding, document OCR, and visual agent capabilities continue to improve steadily.

MiMo-V2.5

MiMo-V2.5 is Xiaomi's open multimodal LLM released in April 2026. Built on a sparse MoE architecture with 310B total and 15B active parameters, it offers a 1M context window plus native visual and audio understanding. It scores 62.3 on the Claw-Eval general subset, matches Gemini 3 Pro on video understanding, and is on par with Claude Sonnet 4.6 on multimodal agentic work. Weights and the model card are available on Hugging Face.

ChatGPT Images 2.0

ChatGPT Images 2.0 is OpenAI's newest image generation model, exposed in the API as gpt-image-2, and the company's first image model with built-in reasoning. Text rendering improves markedly, including non-Latin scripts such as Chinese, Japanese, and Korean, alongside stronger instruction following, multiple aspect ratios, and fine details like dense text and UI elements at up to 2K resolution in the API. It is available now to all ChatGPT and Codex users, with reasoning-enabled outputs offered to Plus, Pro, and Business subscribers.

HappyHorse-1.0-I2V

HappyHorse-1.0-I2V is an image-to-video model on Alibaba Cloud Model Studio. It generates physically realistic, motion-smooth video from a single first-frame image, with optional text prompts for guidance. Output covers 480P, 720P, and 1080P at 24fps MP4; clips run 3 to 15 seconds, and the aspect ratio follows the input image.

HappyHorse-1.0-T2V

HappyHorse-1.0-T2V is a text-to-video model offered on Alibaba Cloud's Bailian platform. It accurately interprets text prompts and produces high-quality videos with realistic physics and fluid motion. Each generation outputs a 3 to 15 second MP4 clip with audio, available in 720P or 1080P at 24 fps, with multiple aspect ratios such as 16:9 and 9:16 for short-form video and creative content production.

Kimi K2.6

Kimi K2.6 is an open-source, natively multimodal agentic model from Moonshot AI. Built on a 1T-parameter MoE architecture with 32B activated parameters per token and a 256K context window, it accepts both image and video input. It targets long-horizon coding and autonomous execution, scaling to 300 coordinated sub-agents, and outperforms GPT-5.4 and Claude Opus 4.6 on benchmarks such as SWE-Bench Pro, DeepSearchQA, and tool-augmented HLE-Full. Weights and code are released under the Modified MIT License, deployable with vLLM and SGLang.

grok-4.3

grok-4.3 is a text LLM available through the xAI API, described by xAI at the time as the fastest and most intelligent model it had built, topping leaderboards in agentic tool calling and instruction following. It offers a 1M token context window, four reasoning effort levels (none, low, medium, and high), and costs $1.25 per 1M input and $2.50 per 1M output tokens. It also serves as the official replacement for retired slugs such as grok-3 and the Grok 4 Fast family.

Qwen3.6-35B-A3B

Qwen3.6-35B-A3B is the first open-weight model in Alibaba's Qwen3.6 series, a native vision-language model built on a hybrid architecture that combines linear attention with a sparse MoE, totaling 35B parameters with 3B activated per token. It provides a native context window of 262,144 tokens, extensible to roughly 1M. Compared with Qwen3.5-35B-A3B, it improves agentic coding, math and code reasoning, spatial intelligence, and object grounding and detection, and is released under the Apache-2.0 license.

Qwen3.6-Flash

Qwen3.6-Flash is the Flash-tier model in Alibaba Cloud's Qwen3.6 native vision-language series on Bailian, accepting image, text, and video input with a 1M-token context window. It improves clearly over 3.5-Flash, with focused gains in agentic coding, where it surpasses the previous generation on multiple coding agent benchmarks, along with math and code reasoning. Vision capabilities also advance in spatial intelligence, object localization, and detection. It supports Function Calling and structured output, with input priced from CNY 1.2 per million tokens in the Beijing region.

Qwen3.6-Flash-2026-04-16

Qwen3.6-Flash-2026-04-16 is Alibaba's fast-tier entry in the Qwen3.6 native vision-language family, covering text generation, deep thinking, and visual understanding, with solid gains over 3.5-Flash. It prioritizes agentic coding, substantially outperforming its predecessor across multiple coding agent benchmarks, and improves both math and code reasoning. Visual spatial intelligence is notably stronger, especially in object localization and detection. The dated snapshot suits production workloads that need stable model behavior.

Claude Opus 4.7

Claude Opus 4.7 is Anthropic's latest Opus model, released April 16, 2026 as a direct upgrade to Opus 4.6 focused on advanced software engineering and long-running agentic tasks. It clears 70% on CursorBench versus 58% for Opus 4.6, and resolves 3x more production tasks on Rakuten-SWE-Bench. Image input now supports up to 2,576 pixels on the long edge, over three times the pixel count of previous Claude models. It is available across Claude apps, major cloud platforms, and the API at $5/$25 per million input/output tokens.

grok-build-0.1

grok-build-0.1 is xAI's coding model trained specifically for agentic coding tasks, including web development, debugging, and MCP support. Available in public beta through the xAI API, it is the same model that powers Grok Build. It is served at over 100 tokens per second and priced at $1 per million input tokens and $2 per million output tokens, making it a fast, economical option for general-purpose agentic and tool calling workloads beyond coding.

GPT-Rosalind

GPT-Rosalind is OpenAI's frontier reasoning model for life sciences research, covering biology, drug discovery, and translational medicine, with stronger tool use and deeper grounding in chemistry, protein engineering, and genomics. It leads BixBench among models with published scores and outperforms GPT-5.4 on 6 of 11 LABBench2 tasks. It is available as a research preview in ChatGPT, Codex, and the API through a trusted access program.

Gemini 3.1 Flash TTS

Released on April 15, 2026, Gemini 3.1 Flash TTS (gemini-3.1-flash-tts-preview) is Google's affordable and expressive text-to-speech model built for low-latency, controllable speech generation. It handles single- and dual-speaker audio with 30 prebuilt voices, and natural-language prompts steer style, accent, pace, and tone, while expressive audio tags like [whispers] allow fine-grained narration control. With automatic language detection, streaming output, and batch support, it suits podcast and audiobook production.

V8.1

Midjourney V8.1 is the next alpha release in the V8 image generation series, keeping the consistent V7-style aesthetic with more stable moodboards and srefs. HD mode runs 3x faster and 3x cheaper and is now the default, while standard resolution is 50% faster and 25% cheaper. The release also brings back image prompts with image weights, adds a Prompt Shortener, and updates Describe to generate longer, more detailed prompts. It is currently available for early testing on alpha.midjourney.com.

Qwen3.6-Max-Preview

Qwen3.6-Max-Preview is Alibaba Cloud's largest and most capable Max model in the Qwen3.6 series, offered as a text-only preview. Compared with Qwen3-Max and Qwen3.6-Plus, it further improves vibe coding and front-end development, runs coding agents more efficiently, and upgrades long-tail knowledge. It provides a 256K context window and supports thinking mode and Function Calling.