Claude Fable

Claude Fable 5.2 is listed on OpenRouter.

• First listed 2026-11-07 as anthropic/claude-fable-5.2.
• Platform listing, not a lab announcement. See the source for availability and details.

Lucy 6 is listed on fal.ai.

• First listed 2026-11-05 as decart/lucy-2-6/realtime.
• Platform listing, not a lab announcement. See the source for availability and details.

LongCat 2.5 is listed on Hugging Face.

• First listed 2026-11-04 as meituan-longcat/LongCat-2.5.
• Platform listing, not a lab announcement. See the source for availability and details.

Genie 4: A new generation of world simulation

• Official model announcement. See the source for capabilities and availability.

GPT-6.1 Terra brings fast and capable computer use to everyone.

• OpenAI summary: "Bringing fast and capable computer use to everyone."
• Official model announcement. See the source for capabilities and availability.

Baidu releases Unlimited-OCR 2 open weights under MIT license on Hugging Face.

• 3.34B vision-language OCR checkpoint (arXiv:2610.39308) with 147,000+ downloads in its first week.
• Handles dense multi-column document parsing, mathematical formulas, and handwriting recognition.

Kimi K3.1 Pro and Highspeed launch across Kimi apps, Code, and API.

• Post-trained from K3 with improvements to coding, 3D creation, and agentic workflows, according to Moonshot.
• One-million-token context; Highspeed is advertised as 2–3 times faster.

Voxtral 3 brings new speech-to-text and text-to-speech models and a voice-agent SDK.

• Voxtral Small Transcribe 3, Mini Transcribe 3 (4B), Mini Transcribe Realtime 3 and Mini TTS 3.
• Voice-agent pipeline combines listening, reasoning, tool calls and speech in one stream.

DeepSeek V4.1-Flash-1017 lands with integrated Search & Fetch APIs and sub-second reasoning.

• Drastically reduces agent memory overhead while maintaining top-tier math and coding benchmarks.
• Full open weights available under MIT license.

MiniMax releases MiniMax-M3.1-Flash multimodal open weights on Hugging Face.

• 220B-parameter vision-language MoE optimized for fast long-context reasoning.
• Ships as open safetensors weights on Hugging Face.

Fish Audio drops Drama 3: 35B MoE emotive speech model with 16-character multi-speaker synthesis.

• Directable performance via natural language bracket tags ([whispers], [laughs], [anger], [crying]).
• Multi-character dialogue handling up to 16 distinct speaker tags in a single request.

Claude Cyber Shield 2.0 deploys real-time continuous supply-chain audit and dependency invariant verification.

• Proactively scans open source registries for behavioral anomalies and malicious code injection.
• Generates automated, cryptographically verifiable remediation PRs.

Kling 4.0 launches globally with native 4K 10-bit HDR and Omni Reference.

• Single-pass video generation expanded up to 30 continuous seconds.
• Omni Reference combines up to 15 reference assets and 10 keyframe controls.

MiniMax launches H3 Max Director mode with native camera path scripting and 3D relighting.

• Enables virtual cinematography with pan, zoom, roll, and dolly trajectory controls.
• Dynamic relighting tools allow shifting time-of-day lighting post-generation.

Big Sleep 2.0 discovers autonomous zero-days in kernel subsystems and opens automated patch validation.

• First AI agent capable of end-to-end vulnerability reproduction, triage, and verifiable patch drafting.
• Expanded from single user-space libraries to complex operating system kernel modules.

Anthropic ships Claude Code 1.2 with automated cross-repository refactoring and git verification.

• Direct terminal agent runs builds, tests, git commits, and interactive PR drafting.
• Built-in sandbox isolates bash execution and verifies tests before finishing.

Claude Haiku 5.5 ships with 1M context window and near-Opus coding speed.

• Expands context to 1,000,000 tokens while maintaining sub-second response times.
• Adaptive thinking mechanism scales reasoning compute dynamically.

Ling 3.1 Flash expands reasoning performance and instruction adherence.

• repo inclusionAI/Ling-3.1-flash (created 2026-10-13, modified 2026-10-15)
• Source adapter: hf-api

Gemini 4 Argon opens to developers, enterprises and consumers.

• Available via the API in AI Studio and the Gemini platform, and in the Gemini app for Google AI Ultra and Pro subscribers.
• Introductory price of $2 per million input tokens and $10 per million output tokens.

Meta introduces Muse Wonder and Muse Wonder Full, frontier models for agents.

• Meta calls them its most capable models to date.
• Active Memory carries context across conversations; Active Focus keeps long tasks on track.

Skywork releases Matrix-Game 4.0 interactive world model weights under Apache-2.0.

• Controllable video-to-video and action-conditioned world generation on Hugging Face.
• Apache-2.0 open-weight release from Skywork AI.

Cartesia Sonic 2 sets real-time voice benchmark with sub-90ms full-duplex conversational synthesis.

• Engineered specifically for low-latency agent architectures.
• Emotion control slider via WebSocket API.

Suno v4.5 introduces stem-level studio mixing, custom lyrical structure controls, and lossless 48kHz audio.

• Allows downloading separated stems (vocals, drums, bass, synth) directly from any generated track.
• Enhanced genre blending and acoustic instrument warmth.

Tencent open-sources Hunyuan3D 2.0 with production-ready PBR material generation.

• Outputs clean topology ready for game engines (Unreal Engine 5, Unity).
• Fast two-stage diffusion generates textured mesh in under 15 seconds.

AlphaFold 3.2 expands de novo molecular drug design with quantitative binding affinity scoring.

• Directly generates novel molecular binders targeting previously undruggable viral proteins.
• Integrated structural confidence scoring for RNA-protein complex interactions.

Google DeepMind unveils Genie 2, generating endless interactive playable 3D worlds in real time.

• Takes keyboard/controller inputs and generates consistent 3D world physics on the fly.
• Capable of simulating complex platformers, racing, and first-person perspectives.

Qwen-Image 2.1 Turbo lands on Hugging Face for few-step image synthesis.

• 7.1B distilled checkpoint accelerating Qwen-Image 2.1 generation.
• Accompanied by Prompt-Editing (PE) text-to-image and image-to-image variants.

Moshi 1.5 drops latency to 110ms with multilingual conversational turn-taking and emotional mirroring.

• Native speech-to-speech dialogue with instantaneous human interruption handling.
• Open weights and Mimi audio codec updated with enhanced voice timbre clarity.

Oasis 1.5 reaches 60 FPS real-time neural simulation with multi-agent persistent memory.

• Fluid gameplay with zero game engine code; all rendering and world mechanics emerge from weights.
• Persistent terrain changes survive long-horizon exploration sessions.

FLUX.2 Turbo drops latency to 8 steps with complete typography prompt adherence.

• Ultra-fast distilled weights for real-time graphics pipelines.
• Renders multi-line typography with zero spelling distortion.

Mistral Large 4: a 1-trillion-parameter natively multimodal MoE.

• 52 billion active parameters.
• Public preview API live in Mistral Studio, served from European datacenters.

Fish Speech 2.0 introduces sub-100ms streaming TTS.

• Zero-shot voice cloning from 3 seconds of audio.

Midjourney V8.5 introduces direct web-canvas camera re-framing and 4K photorealism.

• Interactive 3D bounding box controls for multi-character scene consistency.
• Significant improvement in complex hand anatomy and legible signage text.

Stable Audio 2.5 expands diffusion style-transfer to full 5-minute continuous audio arrangements.

• Multi-track stems export with discrete drum and bass separation.
• Trained exclusively on licensed commercial music catalogs.

Ideogram 4.5 is listed on fal.ai.

• First listed 2026-09-30 as ideogram/v4.5/edit, ideogram/v4.5.
• Platform listing, not a lab announcement. See the source for availability and details.

Gemini 4 Argon announced, rolling out first to trusted cyber defenders.

• Early access runs through Google’s Fairwind Program while safeguards are tested before broad availability.

Claude Opus 5.5 takes #1 on frontier reasoning and autonomous software engineering.

• Sets state-of-the-art across SWE-bench Verified and competitive coding benchmarks.
• Native multi-agent orchestration for continuous 24-hour development runs.

GPT-6.1 Sol offers near-Astra intelligence at one-fifth of Astra’s API token prices.

• OpenAI summary: near-Astra intelligence for coding, computer use and professional work.
• Priced at one-fifth of Astra’s standard API input and output token prices.

Eleven v4, ElevenLabs’ most emotive model, launches with a new architecture.

• Adds a Turbo model for low-latency use cases.
• ElevenLabs cites more authentic, consistent voice cloning across languages.

Claude Sonnet 5.5 is listed on OpenRouter and models.dev.

• First listed 2026-09-28 as anthropic/claude-sonnet-5.5, anthropic/claude-sonnet-5.5:batch.
• Platform listing, not a lab announcement. See the source for availability and details.

GPT-6 Sol and Luna bring frontier intelligence to everyday work.

• Two models with different balances of capability and cost.
• Official model announcement. See the source for capabilities and availability.

Grok 4.7 is listed on OpenRouter and models.dev.

• First listed 2026-09-21 as x-ai/grok-4.7.
• Platform listing, not a lab announcement. See the source for availability and details.

Veo 3.2 integrates Google DeepMind spatial audio and 4K visual resolution.

• Cinematic lighting controls and temporal consistency across scene transitions.
• Deep integration with YouTube Shorts and Gemini creator workflows.

ElevenMusic 2.5 is listed on fal.ai.

• First listed 2026-09-14 as elevenlabs/music/v2.5.
• Platform listing, not a lab announcement. See the source for availability and details.

Alibaba releases Qwen-Image 2.1 open weights on Hugging Face.

• 7.1B-parameter text-to-image foundation model with over 220,000 Hub downloads.
• Supports bilingual typography rendering and structured prompt adherence.

Fable 5.1 expands multi-agent alignment and complex scientific theorem proving.

• Pinnacle intelligence tier exceeding Opus in sustained mathematical rigor.
• Strict safety evaluations and verified cyber-safeguards.

Qwen3.8 Max 0902 checkpoint introduces refined multi-turn tool-use.

• openrouter lists qwen/qwen3.8-max-0902 ("Qwen: Qwen3.8 Max (0902)") from 2026-09-03
• Source adapter: openrouter

GPT-6 Astra debuts as OpenAI’s most intelligent and aligned model yet.

• OpenAI cites state-of-the-art capabilities across computer use, coding, cybersecurity and science.

Gemini 3.8 Flash and 3.8 Flash Cyber arrive three weeks after 3.7.

• Google calls 3.8 its best reasoning and coding model yet, at the speed and cost of 3.7.
• 3.8 Flash Cyber is a cybersecurity variant.

Muse Spark 1.3 is listed on OpenRouter and models.dev.

• First listed 2026-09-02 as meta/muse-spark-1.3.
• Platform listing, not a lab announcement. See the source for availability and details.

Z.ai releases GLM-5.3 and GLM-5.3-Flash open-weight checkpoints on Hugging Face.

• 753B-parameter MoE flagship alongside a high-throughput Flash variant.
• Over 6.3 million downloads across the GLM-5.3 release repositories on Hugging Face.

DeepSeek R1 matches frontier reasoning with open weights and emergent self-verification chains.

• Trained via large-scale reinforcement learning without cold-start SFT.
• Distilled dense variants run on single consumer workstations.

Gemini 3.7 Flash, Google’s most intelligent workhorse model yet for coding and agents.

• Official model announcement. See the source for capabilities and availability.

Grok 4.6 is listed on OpenRouter and models.dev.

• First listed 2026-08-12 as x-ai/grok-4.6.
• Platform listing, not a lab announcement. See the source for availability and details.

LTX 2.5 is listed on fal.ai.

• First listed 2026-08-11 as lightricks/ltx-2.5/image-to-video/pro, lightricks/ltx-2.5/image-to-video/fast, lightricks/ltx-2.5/text-to-video/pro.
• Platform listing, not a lab announcement. See the source for availability and details.

MAI-Image-2.6 launches at No. 2 on Arena’s text-to-image leaderboard.

• Microsoft reports +79 Elo over MAI-Image-2.5 overall and +91 Elo on text rendering.
• MAI-Image-2.6 and 2.6-Flash entered Public Preview on Microsoft Foundry (post updated 4 September 2026).

Seedream 5.0 Pro adds grounded, layer-separated editing.

• Edits a single region while holding the rest of the frame.
• Layer separation for compositing.

Tripo3D image-to-3D object generation.

• Single image in, 3D model out.

Muse Spark 1.2 ships with Muse Code, a terminal coding agent.

• Terminal-Bench 2.1 up to 82.9% from 76.2%; DeepSWE v1.1 to 59.3% from 53.0%.
• Trained on whole software engineering workflows rather than completion.

Qwen3.8-Max, 2.4T parameters and the first Max tier to open.

• 2.4 trillion parameters, 95B active. Text, image and video, 1M context.
• First Max-class Qwen to be open-sourced — a return to the open strategy after several proprietary flagships.

DeepSeek V4 Flash delivers high-throughput reasoning at record-low inference cost.

• Flagship speed tier for production agent deployments.

MiniMax H3 text-to-video lands with high frame consistency and fluid physics.

• First generation H3 model release.

Hi3D generates 3D models from a single image.

• Direct competitor to Tripo in the image-to-3D slot.

Full K3 weights published on Hugging Face, on the date Moonshot committed to.

• Moonshot calls K3 the world’s first open 3T-class model.
• Released under Moonshot’s own kimi-k3 licence.

Qwen3.7-Flash lands between the 3.5 train and 3.8-Max.

• Flash tier; announcement date not yet confirmed against Alibaba.

Opus 5 lands below Fable 5, at half the price.

• New Opus-tier flagship above Opus 4.8.
• Anthropic positions it near Fable 5's intelligence at $5/Mtok input against Fable 5's $10.

Ling-3.0-flash released with open weights.

• Throughput-oriented MoE.
• Part of the July open-weight wave.

MAI-Image-2.5-Pro introduced alongside MAI-Voice-2-Flash.

• Microsoft highlights accurate text rendering and natural-language edits.
• Also listed on fal.ai as microsoft/mai-image-2.5-pro.

Three Qwen releases inside 72 hours.

• Rapid-fire release train across sizes.
• Open weights on all three.

Laguna S 2.1 released into the July model wave.

• Incremental gains on repo-scale editing.

GPT-5.6 arrives as a lineup: Sol, Terra and Luna.

• Three named variants released together instead of a single flagship.
• Splits reasoning depth, speed and cost across the tiers.

Sonilo V1.1 generates licensed music from one prompt.

• Control over style, mood, instrumentation and exact duration.
• Licensing is the differentiator, not fidelity.

Gemini 3.6 Flash ships as a three-model trio.

• Three Flash variants released simultaneously.
• Aimed at high-volume, latency-sensitive workloads.

LongCat 2.0, a 1.6T-parameter sparse MoE from Meituan.

• 48B active parameters.
• Positioned for coding and repository tasks.

Grok 4.5 pushes coding and knowledge work at lower token use.

• Headline claim is efficiency, not raw capability.
• Fewer tokens consumed per completed task.

Muse Spark 1.1 leans hard into computer use.

• 1,000,000-token context window.
• Positioned on agent work rather than chat.

Kimi K3 launches at 2.8T parameters with native vision and a 1M-token context.

• Built on Kimi Delta Attention and Attention Residuals, per Moonshot.
• Available on Kimi.com, Kimi Work, Kimi Code and the Kimi API at launch.

LTX-2.3 ships text-, image- and audio-to-video.

• Pro and Fast variants.
• Generates synchronised audio with video.

FLUX 3 enters early access — video, audio and robot actions from one model.

• First BFL model to generate video, audio and robot action-prediction from a single set of weights.
• FLUX 3 Video does up to 20-second clips with native audio.

Lucy 2.5 edits webcam and streamed video live.

• Restyle, background swap and object replacement at interactive rates.
• Runs over WebRTC rather than as a batch job.

Reve 2.1 edits from text prompts with layout intelligence.

• Prompt adherence and accurate text rendering.

KAT-Coder V2.5 lands in Pro and Air sizes.

• Both variants shipped the same day.
• Scoped at whole issues and workflows.

Seedream 5.0 Lite editing endpoint.

• Fast variant, multi-input editing.

Hy3 ships with configurable reasoning.

• 21B active of 295B total.
• 192 experts with top-8 routing.

Ideogram V4.0 sharpens text rendering and fine detail.

• Accurate typography remains the selling point.
• Posters and logos as the target use.

Sonnet 5 arrives under the Opus 5 flagship.

• Tracked separately from Opus — different cadence, different tier.

Nano Banana 2 Lite, also listed as Gemini 3.1 Flash Lite Image.

• Same model under two ids — the case the alias map exists for.
• Lite tier below Nano Banana 2.

Baidu open-sources Unlimited-OCR with 3.34B MIT-licensed weights on Hugging Face.

• Over 4.2 million Hub downloads and 4,900+ likes since release (arXiv:2606.23050).
• End-to-end image-text-to-text document recognition in safetensors format.

Seedance 2.5 extends clip length and shot consistency.

• Now the strongest long-form image-to-video option.
• Better identity retention across cuts.

GLM-5.2 ships 753B MIT-licensed open weights on Hugging Face.

• MIT-licensed open-weight foundation model for reasoning and agentic coding.
• Published in BF16 and FP8 weight configurations.

Fable 5 ships as Anthropics premier Mythos-class intelligence engine.

• Designed specifically for multi-day uninterrupted agentic task planning.
• Introduced prior to the Opus 5 release window.

Ray3.2 adds frame-by-frame direction and a public API.

• Clips up to 20 seconds at 1080p.
• Up to 16 keyframes per clip to shape motion and pacing.

TripoSplat converts one image into 3D Gaussians.

• Open-source, from TripoAI / VAST AI Research.

MAI-Image-2.5 and 2.5-Flash launch; 2.5 ranks No. 2 on Arena’s image-edit leaderboard.

• Generation and precise, localized editing in one model.
• 2.5-Flash targets fast, scalable production workloads.

MiniMax-M3 ships 427B multimodal MoE weights on Hugging Face.

• 427B-parameter native image-text-to-text architecture.
• Released with BF16 and MXFP8 weight checkpoints.

PaddleOCR-VL 1.6 upgrades compact 0.96B multilingual document and table parsing on Hugging Face.

• Apache-2.0 open weights supporting vLLM and Transformers inference for text, tables, formulas, and charts.
• 87,000+ Hub downloads with ultra-low sub-1B parameter footprint.

First Muse Spark release.

• Agent-first framing from launch.

Recraft V4.1 brings quieter photorealism and new illustration styles.

• Adds a Utility variant for when simplicity matters more than creative surprise.

V8.1 is the fastest Midjourney yet — 4–5× on standard jobs.

• HD 2K output added.
• Better prompt adherence and small-detail retention.

Qwen 3 generation opens with hybrid reasoning.

• Thinking and non-thinking modes in one checkpoint.

Hy3 preview, ahead of the full release.

• High-efficiency MoE for agentic workflows.

V3 with style training from reference images.

• Fine-tuning on a small reference set.

Kimi K2.6 open-sources long-horizon coding and agent swarms.

• Available via Kimi.com, the Kimi App, the API and Kimi Code.

MAI-Image-2-Efficient: flagship quality at nearly 41% lower cost.

• Microsoft reports 22% faster generation.

Incremental bump to the 4-series flagship.

• Coding and tool-use gains over 4.5.
• Same pricing as 4.5.

MiniMax-M2.7 open-weight 228B MoE lands on Hugging Face.

• 228B-parameter text reasoning checkpoint preceding the M3 generation.

ElevenMusic launches on iOS, unifying voice, music and SFX.

• One app across voice, music and sound effects.
• Positioned against Suno on licensing rather than quality.

Matrix-Game 3.0 open-sources image-and-text-conditioned playable world generation.

• Apache-2.0 open weights for interactive environment synthesis.

v5.5 adds custom fine-tuning and verified voice cloning.

• Better vocal expressiveness over v5.
• Custom model fine-tuning for users.

Sora 2 deprecated; API removal set for 24 September 2026.

• OpenAI notified developers on 24 March 2026.
• The Videos API and every sora-2 and sora-2-pro snapshot leave the API on 24 September 2026.

MAI-Image-2 targets natural light, accurate skin tones and lived-in scenes.

• Official model announcement. See the source for capabilities and availability.

Gemini 3 generation reaches the Flash tier.

• 1M-token context standard across the tier.

Laguna S 2 targets whole-repository reasoning.

• Trained on execution feedback rather than static code alone.

Nano Banana 2, Google's image generation and editing model.

• Generation and editing in one endpoint.

Kling 3.0 adds multilingual lip sync and cinematic motion controls.

• Lip sync across multiple languages in one pass.
• Holds four separate entries in the video top ten.

5.2 tightens routing between fast and thinking modes.

• Less visible mode-switching.
• Lower latency on short turns.

Recraft V4, a ground-up rebuild focused on design taste.

• Targets visual taste, prompt accuracy and output quality for production design work.

Seedance 2.0 takes the #1 slot on Artificial Analysis.

• Displaces Veo and Kling at the top of the video board.
• Large jump in motion coherence.

PaddleOCR-VL 1.5 ships Apache-2.0 vision-language OCR checkpoints on Hugging Face.

• Sub-billion-parameter document layout and OCR model from Baidu's PaddlePaddle team.

Kimi K2.5 turns K2 into a native multimodal model.

• Continued pretraining on about 15T mixed visual and text tokens, per Moonshot.
• Can direct an agent swarm of up to 100 sub-agents.

Ray3.14: native 1080p, 3x cheaper and 4x faster.

• Luma positions it as removing the quality, speed and cost tradeoff of Ray3.

V8 resets the base model.

• Large jump in prompt following.

FLUX klein targets local and enterprise inference.

• 40% less memory on NVIDIA RTX cards.
• Roughly 40% faster.

v2 improves stem separation and instrumental control.

• Preferred by musicians over Suno on instrumental work.

Runway Gen-4.5 tops the Artificial Analysis text-to-video leaderboard.

• Runway cites 1,247 Elo on the leaderboard as of 30 November 2025.

FLUX.2 replaces FLUX.1 as the flagship.

• Leads open-weight image quality.
• Strong on photoreal and commercial work.

Opus 4.5 lands with a large price cut.

• Roughly a third of 4.1 pricing.
• Effort control added to the API.

Veo 3.1 brings true 4K at up to 60fps.

• 3840×2160 output.
• Ambient sound, dialogue and effects generated with the video.

MAI-Image-1 debuts in the top 10 on LMArena.

• Later added to Bing Image Creator.

Sora 2 launches with synced audio and a social app.

• Best-in-class physical motion.
• Shipped with a consumer feed app.

2.5 Turbo cuts cost while improving motion.

• Faster generation at lower price per second.

v5 sets the current quality benchmark at 1,293 ELO.

• Ahead of every competitor on fidelity, structure and vocal realism.

Luma launches Ray3.

• New generation of the Ray video model; Ray2 and Dream Machine are now deprecated.

GPT-5 replaces the 4-series across ChatGPT.

• Unified router over reasoning and non-reasoning paths.

Opus 4.1 refines agentic search and coding.

• Drop-in replacement for Opus 4.

ElevenLabs enters music generation.

• Built on licensed training data from the start.

Wan 2.2 releases open weights and inference code.

• Text-to-video and image-to-video A14B models plus a 5B text-image-to-video model.

K2 established the open MoE line.

• 1T-parameter MoE, agentic tool use.

Grok 4 released with a heavy multi-agent tier.

• Grok 4 Heavy runs multiple agents in parallel.

First public Seedance release.

• Multi-shot generation from a single prompt.

v3 adds audio tags for directed emotional delivery.

• Inline tags control emotion and pacing.
• 70+ languages.

Veo 3 introduces synchronised native audio.

• First major model to generate video and matching audio together.

v4.5 leads on lyric integration.

• Best-in-class lyric-to-melody alignment at release.

Kling 2.0 rebuilds the motion model.

• Large improvement in physical plausibility.

V7 introduces draft mode and personalisation by default.

• Personalisation on by default for the first time.

Runway Gen-4 keeps characters, locations and objects consistent across scenes.

• Official model announcement. See the source for capabilities and availability.

Alibaba open-sources Wan 2.1 matching commercial video generators with consumer GPU support.

• Supports both T2V and I2V at native 1080P resolution.
• 1.3B model runs on consumer RTX 4090 GPUs.

Anthropic deploys specialized Claude cybersecurity defense suite for automated patch verification.

• Provides deep binary and source-level security audits with verifiable proof-of-fix.
• Rigorous safety evaluation preventing dual-use weaponization.

Big Sleep discovers first-ever in-the-wild zero-day in widely used open source software (SQLite).

• Evolved from Project Naptime into a specialized LLM-assisted vulnerability discovery agent.
• Discovered stack buffer underflow in SQLite before it reached any production release.

Decart and Etched release Oasis, a real-time neural game engine running Minecraft without game code.

• Takes keyboard and mouse inputs and generates the next video frame autoregressively.
• Simulates physics, inventory, biome rendering, and ray-traced lighting purely from weights.

Recraft V3 sets a new image-generation standard, led by text rendering.

• Recraft cites first place on the Artificial Analysis text-to-image benchmark.

Mistral Large 2 delivers 123B parameter frontier reasoning with 128k context.

• State-of-the-art multilingual reasoning rivaling GPT-4o and Llama 3.1 405B.
• Engineered to drastically minimize hallucination and sycophancy.

Kyutai open-sources Moshi, achieving conversational speech-to-speech without STT/TTS cascade.

• First native audio-to-audio foundation model with realistic emotional prosody and backchanneling.
• Full duplex: can listen and speak simultaneously without audio cutoffs.

AlphaFold 3 predicts the structure and interactions of all of lifes molecules.

• Goes beyond proteins to model DNA, RNA, ligands, antibodies, and chemical modifications.
• Diffusion-based architecture replacing previous Evoformer.

Stable Audio 2.0 introduces audio-to-audio style transfer and 3-minute stereo generation.

• Generates full-length structured tracks (intro, development, outro) from text or audio prompts.
• Trained on AudioSparx licensed catalog for copyright integrity.