New AI Models

Nemotron-3-Diarization-preview

Nemotron-3-Diarization-preview is an early access speaker diarization model from NVIDIA, built on the streaming sortformer architecture and distributed through the NeMo framework as a .nemo checkpoint. It handles speaker diarization and speaker tagging for multi-speaker audio, and the streaming sortformer design processes live audio in small overlapping chunks so each speaker keeps a consistent label across the stream. Access is gated under the NVIDIA Software and Model Evaluation License for internal testing and evaluation on NVIDIA GPUs, with intended uses covering meeting and broadcast transcription, speech dataset annotation, and TTS training-data annotation.

GPT‑5.6

The GPT‑5.6 model family is OpenAI's latest flagship series, spanning Sol, Terra, and Luna, and is now available in Kiro, a software development agent for AI-native coding. It delivers stronger performance per dollar and on-demand capability for complex tasks, applied to long-running development work grounded in requirements, codebases, and team standards. In testing by OpenAI and AWS, GPT‑5.6 Terra completed Terminal-Bench 2.1 tasks in Kiro at roughly 82% cost reduction.

PixVerse R2

PixVerse R2 is the second-generation real-time audiovisual world model from PixVerse Research, the successor to R1, presented as a unified scaling architecture for streaming worlds. It continues pretraining of a single Omni Causal AR backbone, then distills that same backbone into a real-time generator using DDMD with adversarial regularization, block-sparse attention with over 90% sparsity, and a pyramid ultra-few-step scheme, producing synchronized video and audio driven by live text, multimodal reference, audio, or action (WASD) inputs. Long-horizon stability comes from Sink, Rolling, and Object KV memory channels plus an Error Bank that replays failure states during training; internal evaluations report long-run brightness drift reduced from 0.201 to 0.129, a 35.8% reduction. The post is a research announcement and does not address public availability.

DeepSeek-V4-Flash-Vision-Exp

DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal model released on the DeepSeek API Platform on August 21, 2026. It matches DeepSeek-V4-Flash on text tasks, including agents, reasoning, and world knowledge, while showing a major step up on multimodal agent benchmarks that brings it close to Opus-4.8. Images are billed at up to 384 tokens each at V4-Flash pricing and can be passed via base64, external URLs, or the Files API, making it a fit for agent workflows that pair visual understanding with tool calls.

Eleven v3 Conversational

Eleven v3 Conversational is ElevenLabs' most expressive realtime speech synthesis model, producing natural, emotionally rich speech at low latency of roughly 280ms. It supports 70+ languages and provides audio tags for fine-grained control over delivery. Typical uses include realtime support voice agents, AI assistants, and interactive characters, accessed via the new Text to Dialogue WebSocket.

FLUX Video Upscale

FLUX Video Upscale is a standalone video enhancement tool from Black Forest Labs, available through the BFL API and playground. It regenerates video from 480p upward at up to native 4K, with upscale factors of 1.5x, 2x, and 3x. A Precise mode (4 steps, $0.07 per megapixel-second) preserves identity and reference details, while a Creative mode (8 steps, $0.10 per megapixel-second) pushes repair and detail generation further. It fixes smudged faces and gridded artifacts on textures like water and grass, making generated footage delivery-ready for broadcast and large-screen formats.

Wan3.0-Video-Prime

Wan3.0-Video-Prime is the accelerated version of Alibaba Cloud's Wan video generation model, delivering significantly faster generation speed while maintaining high-quality output. It accepts text, image, video, and audio inputs in a single model, covering text-to-video, image-to-video (first frame and first-last frame), and reference-based video generation. It outputs video up to 30 seconds long at 480P, 720P, or 1080P with a 30fps frame rate.

Runway Ruby

Runway Ruby is an SDR to HDR video processing model available in the Runway API under the model name ruby. It converts any SDR video, uploaded or AI-generated, into true HDR, preserving source pixels and audio via a bounded, color-preserving grade rather than re-rendering. Outputs include 10-bit HEVC in HDR10 or HLG, a BT.2020 PQ ProRes .mov mezzanine (422, 422 HQ, 4444), and half-float OpenEXR sequences. Clips up to 30 seconds and 4096 pixels per side are supported, accessed through the /v1/video_to_hdr endpoint for streaming and post-production HDR delivery.

GPT-5.6 Luna

GPT-5.6 Luna is the fastest, most cost-efficient tier in OpenAI's GPT-5.6 lineup, alongside the flagship Sol and the balanced Terra. It approaches GPT-5.5's overall capability at less than half the estimated cost and outscores Claude Opus 4.8 on the Artificial Analysis Coding Agent Index in about a third of the time. Priced at $1 per million input tokens and $6 per million output tokens, with an 80% price cut announced on July 30, it targets cost-sensitive workloads.

GLM-5.3

GLM-5.3 is Zhipu AI's flagship coding model, sharing the GLM-5.2 base model, with all gains from scaled post-training. Z.ai calls it the most capable open-weights model for coding, reaching open-source SOTA on Terminal Bench 3.0 and Agents' Last Exam and improving 50% over GLM-5.2 on the in-house Z.ai Code Bench. Its emergent cyber capability posts a state-of-the-art 84.5% on CyberGym for vulnerability discovery. Weights follow two weeks after launch, and it fits agentic coding and long-horizon work in agents like ZCode and Claude Code.

Qwen3.8-27B

Qwen3.8-27B is a 27B dense vision-language model in Alibaba's Qwen3.8 open-model family, with native image and video understanding. It strengthens coding and office-productivity capabilities in both text and vision modalities over Qwen3.6-27B, posting large gains on benchmarks such as SWE-bench Pro and OSWorld. The model offers a native 256K context window, extensible to 1M tokens, and is suited to coding, professional work, and long-horizon agentic tasks.

Nemotron-Labs-Teacher

Nemotron-Labs-Teacher is a series of open-weight distillation teacher models from NVIDIA, released on Hugging Face on August 14, 2026 in five domain specializations: General-Reasoning, Competition-Coding, Instruction-Following, STEM, and Chat. Each checkpoint starts from the post-trained Nemotron 3 Ultra (550B total, 55B active parameters, LatentMoE architecture, up to 1M context) and adds a round of reasoning-intensive fine-tuning and reinforcement learning. Licensed under OpenMDW-1.1 for commercial and non-commercial use, they serve as domain-specialized teachers for Multi-Teacher On-Policy Distillation (MOPD) and as generators of long-horizon reasoning traces.

DeepSeek-V4-Pro-0813

DeepSeek-V4-Pro-0813 is the general-availability release of DeepSeek's V4 Pro text model, bringing major Agent upgrades with strong production gains. It supports three reasoning effort levels (low, high, max), native OpenAI Responses API compatibility, and one-click setup for Codex. Agentic and coding benchmarks such as Terminal Bench and DeepSWE show clear gains over the V4-Pro preview, and API pricing now uses peak and off-peak rates, with off-peak at half the peak price.

Gemini 3.7 Flash

Gemini 3.7 Flash is Google's workhorse model for coding and agents, released on August 13, 2026, three weeks after Gemini 3.6 Flash. It posts clear gains over its predecessor on benchmarks such as FrontierCode 1.1 Main (43.6% vs 34.4%) and DeepSWE v1.1 (65.3% vs 49.0%), and can build web UI that matches reference screenshots or design systems. Introductory pricing runs $0.75 per million input tokens and $3.75 per million output tokens, and it is available through the Gemini API, Google Antigravity, and AI Studio.

MAI-Cyber-1-Flash

MAI-Cyber-1-Flash is Microsoft's first cybersecurity model, a compact, code-heavy model from the in-house MAI-Thinking-1 lineage, launched on August 13, 2026 inside MDASH, its multi-agent vulnerability identification and remediation harness. It scores 96% on CyberGym, ahead of Mythos, and delivers 50% cost savings over Microsoft's best prior MDASH stack, handling up to 90% of tasks. Aimed at finding and fixing vulnerabilities in complex codebases and agentic code scanning in security operations, it ships as an enterprise offering with tenant isolation and sandboxed execution, plus a published technical report, not open weights.

sign-language-to-text (SL2T)

SL2T is a massively multilingual sign-language-to-text translation model from Google DeepMind and Android that converts signing into streaming text, so users can sign anywhere they would normally type. Trained on more than 100,000 hours of data spanning over 50 sign languages, it launches with American Sign Language (ASL) to English and scores 70 BLEURT zero-shot on the FLEURS-ASL benchmark. It is available at no cost in Gboard and Live Transcribe on Pixel 11, with more devices and languages to follow.

Qwen3.8-2.4T-A95B

Qwen3.8-2.4T-A95B is the open-source release of Qwen's latest flagship series, launched in August 2026. It uses a sparse MoE architecture with 2.4 trillion total parameters and about 95 billion active parameters per step, combining hybrid attention with a 1 million token context window. Key benchmark results include GPQA Diamond 92.6, PaperBench 93.0, OSWorld 86.1, and BabyVision 82.0, with a fourth-place global ranking on CodeArena.

MAI-Code-1.1-Flash

MAI-Code-1.1-Flash is a small, efficient coding model from Microsoft, now in production in GitHub Copilot. Compared with the 1.0 release announced at Microsoft Build in June, it produces higher quality code at 25% greater token efficiency and one quarter of the cost, with a 22% gain on Terminal-Bench 2.1 and 15% on .NET tasks. Tokens also stream 25% faster, making it a practical fit for everyday coding assistance across IDE and CLI workflows.

Nemotron 3.5 Lightning (30B A3B)

NVIDIA Nemotron 3.5 Lightning is an open mixture-of-experts model with 30B total parameters (3B active, A3B) built for long-running agentic AI workloads. NVIDIA reports up to 4x faster output speed and 30% faster agentic task completion than other models in its class, with frontier-level accuracy on PinchBench. Offered with NVFP4 quantization, it runs locally on RTX PCs, DGX Spark, DGX Station and Jetson, and can be post-trained with NeMo for high-volume tasks such as code review and tool use.

Muse Glimmer-30B

Muse Glimmer-30B is a 30-billion-parameter agentic model from Meta Superintelligence Labs, open sourced under Apache 2.0 and optimized for always-on local agent workflows on a single consumer GPU. It handles precise tool calling, long-horizon multi-step reasoning, and interleaved text and image input across more than 100 languages. Meta reports strong results for its size class on agentic benchmarks such as SWE-Bench and tau-Bench versus Gemma4-31B and Qwen3.6-27B, with quantization bringing the model under 20 GB for local coding, function calling, and LLM-as-a-judge evaluation.