MODELS · 99 stories

Models.

Frontier model releases, benchmark results, and the research advancing what AI systems can do.

99

Stories in section

Time:

Sort

Google tests new Gemini Flash build on LM Arena — featured image for AI Insiders

Models / 10 min ago

Google tests new Gemini Flash build on LM Arena

By Alessandro Benigni

$GPT-5.5 Pro and Claude Opus 4.7 paired up to crack open math problems — featured image for AI Insiders$

Models / yesterday

GPT-5.5 Pro and Claude Opus 4.7 paired up to crack open math problems

By Alessandro Benigni

Meituan's LongCat-2.0 outs itself as OpenRouter's stealth hit — featured image for AI Insiders

Models / yesterday

Meituan's LongCat-2.0 outs itself as OpenRouter's stealth hit

By Alessandro Benigni

Anthropic prices Claude Sonnet 5 to undercut its own Opus tier — featured image for AI Insiders

Models / yesterday

Anthropic prices Claude Sonnet 5 to undercut its own Opus tier

By Alessandro Benigni

AI2's DiScoFormer cuts density error 37x over KDE at 100 dimensions — featured image for AI Insiders

Models / yesterday

AI2's DiScoFormer cuts density error 37x over KDE at 100 dimensions

By Alessandro Benigni

Google Cloud will sell SandboxAQ's science models alongside Gemini — featured image for AI Insiders

Models / yesterday

Google Cloud will sell SandboxAQ's science models alongside Gemini

By Alessandro Benigni

Google opens Gemini's personalized image generation to all US free users — featured image for AI Insiders

Models / yesterday

Google opens Gemini's personalized image generation to all US free users

By Alessandro Benigni

Sakana AI ships Fugu Ultra at 93.2 on LiveCodeBench after losing Claude — featured image for AI Insiders

Models / yesterday

Sakana AI ships Fugu Ultra at 93.2 on LiveCodeBench after losing Claude

By Alessandro Benigni

Google retrofits multi-token prediction onto frozen Gemini Nano — featured image for AI Insiders

Models / 2 days ago

Google retrofits multi-token prediction onto frozen Gemini Nano

By Alessandro Benigni

Grok 4.5 enters private beta at SpaceX and Tesla on a 1.5T-parameter base — featured image for AI Insiders

Models / 2 days ago

Grok 4.5 enters private beta at SpaceX and Tesla on a 1.5T-parameter base

By Alessandro Benigni

OpenAI previews GPT-5.6 under U.S. government watch — featured image for AI Insiders

Models / 2 days ago

OpenAI previews GPT-5.6 under U.S. government watch

By Alessandro Benigni

Reward Models Are Too Sensitive, and That Is Why Your RLHF Breaks — featured image for AI Insiders

Models / 2 days ago

Reward Models Are Too Sensitive, and That Is Why Your RLHF Breaks

By Alessandro Benigni

DeepReinforce ships Ornith-1.0, a coding model that writes its own RL scaffolds — featured image for AI Insiders

Models / 5 days ago

DeepReinforce ships Ornith-1.0, a coding model that writes its own RL scaffolds

By Alessandro Benigni

Goodfire removed a model's German fluency by tuning 4 tokens — featured image for AI Insiders

Models / 5 days ago

Goodfire removed a model's German fluency by tuning 4 tokens

By Alessandro Benigni

Liquid AI shrinks its non-transformer model to 230M parameters — featured image for AI Insiders

Models / 5 days ago

Liquid AI shrinks its non-transformer model to 230M parameters

By Alessandro Benigni

GLM-5.2 is the first open model that holds up in a coding agent harness — featured image for AI Insiders

Models / 6 days ago

GLM-5.2 is the first open model that holds up in a coding agent harness

By Alessandro Benigni

OpenAI and Broadcom build Jalapeño, a custom LLM inference chip — featured image for AI Insiders

Models / 6 days ago

OpenAI and Broadcom build Jalapeño, a custom LLM inference chip

By Alessandro Benigni

OpenAI Quietly Upgrades GPT-5.5 Instant Inside ChatGPT — featured image for AI Insiders

Models / 6 days ago

OpenAI Quietly Upgrades GPT-5.5 Instant Inside ChatGPT

By Alessandro Benigni

Baidu's Unlimited OCR parses dozens of pages in a single forward pass — featured image for AI Insiders

Models / Jun 25

Baidu's Unlimited OCR parses dozens of pages in a single forward pass

By Alessandro Benigni

ByteDance's Seedance 2.5 Generates 30-Second 4K Video From One Prompt — featured image for AI Insiders

Models / Jun 25

ByteDance's Seedance 2.5 Generates 30-Second 4K Video From One Prompt

By Alessandro Benigni

Krea releases open-weight image models built to escape default aesthetics — featured image for AI Insiders

Models / Jun 25

Krea releases open-weight image models built to escape default aesthetics

By Alessandro Benigni

A 0.22B inpainting model that matches an 11.9B generalist — featured image for AI Insiders

Models / Jun 24

A 0.22B inpainting model that matches an 11.9B generalist

By Alessandro Benigni

Alibaba's HappyHorse 1.1 passes Sora in video rankings — featured image for AI Insiders

Models / Jun 24

Alibaba's HappyHorse 1.1 passes Sora in video rankings

By Alessandro Benigni

The transformer monoculture is over. Here is what replaced it. — featured image for AI Insiders

Models / Jun 22

The transformer monoculture is over. Here is what replaced it.

By Alessandro Benigni

Diffusion LLMs Are Not an Interpretability Dead End — featured image for AI Insiders

Models / Jun 22

Diffusion LLMs Are Not an Interpretability Dead End

By Alessandro Benigni

Mercury 2 hits 1,000 tokens per second. Here is what that buys you. — featured image for AI Insiders

Models / Jun 22

Mercury 2 hits 1,000 tokens per second. Here is what that buys you.

By Alessandro Benigni

NVIDIA's ZPPO Fixes the Hardest-Question Problem in RL Post-Training — featured image for AI Insiders

Models / Jun 20

NVIDIA's ZPPO Fixes the Hardest-Question Problem in RL Post-Training

By Alessandro Benigni

OpenAI eyes GPT-5.6 launch next week with 1.5M token context — featured image for AI Insiders

Models / Jun 20

OpenAI eyes GPT-5.6 launch next week with 1.5M token context

By Alessandro Benigni

OpenAI shows RL on beneficial traits generalizes across dozens of benchmarks — featured image for AI Insiders

Models / Jun 20

OpenAI shows RL on beneficial traits generalizes across dozens of benchmarks

By Alessandro Benigni

Ai2 ships MolmoMotion to close robotics' language gap — featured image for AI Insiders

Models / Jun 19

Ai2 ships MolmoMotion to close robotics' language gap

By Alessandro Benigni

Cursor is building a 1.5-trillion-parameter model from scratch — featured image for AI Insiders

Models / Jun 19

Cursor is building a 1.5-trillion-parameter model from scratch

By Alessandro Benigni

Kimi K2.7 Code ran 16x cheaper than Claude Fable 5 in a landing-page test — featured image for AI Insiders

Models / Jun 19

Kimi K2.7 Code ran 16x cheaper than Claude Fable 5 in a landing-page test

By Alessandro Benigni

Mistral Previews a Large Sparse MoE This Summer, Opens July Access — featured image for AI Insiders

Models / Jun 19

Mistral Previews a Large Sparse MoE This Summer, Opens July Access

By Alessandro Benigni

OpenAI's new life-science benchmark hits a 36% ceiling — featured image for AI Insiders

Models / Jun 19

OpenAI's new life-science benchmark hits a 36% ceiling

By Alessandro Benigni

Microsoft Extends Phi Silica to NVIDIA GPUs, Testing NPU Limits — featured image for AI Insiders

Models / Jun 18

Microsoft Extends Phi Silica to NVIDIA GPUs, Testing NPU Limits

By Alessandro Benigni

OpenAI's GPT-Bidi-1 aims to fix voice mode's turn-taking problem — featured image for AI Insiders

Models / Jun 18

OpenAI's GPT-Bidi-1 aims to fix voice mode's turn-taking problem

By Alessandro Benigni

Qwen-RobotWorld makes plain language the control layer for robots — featured image for AI Insiders

Models / Jun 18

Qwen-RobotWorld makes plain language the control layer for robots

By Alessandro Benigni

Weibo's 3B model matches flagships on LeetCode, falls short on science — featured image for AI Insiders

Models / Jun 18

Weibo's 3B model matches flagships on LeetCode, falls short on science

By Alessandro Benigni

Z.ai ships GLM-5.2 with a 1M-token context for full codebases — featured image for AI Insiders

Models / Jun 18

Z.ai ships GLM-5.2 with a 1M-token context for full codebases

By Alessandro Benigni

Moonshot AI ships Kimi K2.7 Code, a 1T-parameter MoE drop-in for OpenAI APIs — featured image for AI Insiders

Models / Jun 16

Moonshot AI ships Kimi K2.7 Code, a 1T-parameter MoE drop-in for OpenAI APIs

By Alessandro Benigni

Z.ai's GLM-5.2 ships with MIT license and 1M-context coding support — featured image for AI Insiders

Models / Jun 16

Z.ai's GLM-5.2 ships with MIT license and 1M-context coding support

By Alessandro Benigni

Google ships DiffusionGemma, a 26B open model that generates text 4x faster — featured image for AI Insiders

Models / Jun 12

Google ships DiffusionGemma, a 26B open model that generates text 4x faster

By Alessandro Benigni

The Fable 5 leak's real story is 120,000 characters — featured image for AI Insiders

Models / Jun 12

The Fable 5 leak's real story is 120,000 characters

By Alessandro Benigni

Anthropic ships Mythos-class capability to the public via Claude Fable 5 — featured image for AI Insiders

Models / Jun 11

Anthropic ships Mythos-class capability to the public via Claude Fable 5

By Alessandro Benigni

Xiaomi hits 1,000 tokens/sec on a trillion-param model using commodity GPUs — featured image for AI Insiders

Models / Jun 10

Xiaomi hits 1,000 tokens/sec on a trillion-param model using commodity GPUs

By Alessandro Benigni

Apple pays Google $1B/yr to admit it lost the AI race — featured image for AI Insiders

Apple pays Google $1B/yr to admit it lost the AI race

By Alessandro Benigni

Claude matches ChemDraw on NMR, then does what ChemDraw cannot — featured image for AI Insiders

Claude matches ChemDraw on NMR, then does what ChemDraw cannot

By Alessandro Benigni

Google ships Gemma 4 QAT models that fit in 1GB on a phone — featured image for AI Insiders

Google ships Gemma 4 QAT models that fit in 1GB on a phone

By Alessandro Benigni

Anthropic's Oceanus red team leaked to Chinese proxy within hours — featured image for AI Insiders

Anthropic's Oceanus red team leaked to Chinese proxy within hours

By Alessandro Benigni

NVIDIA ships a unified safety model with auditable reasoning — featured image for AI Insiders

NVIDIA ships a unified safety model with auditable reasoning

By Alessandro Benigni

Google's Sleep+Dreaming turns idle time into a training loop — featured image for AI Insiders

Google's Sleep+Dreaming turns idle time into a training loop

By Alessandro Benigni

Meta's Muse Spark Has No API Launch Date — featured image for AI Insiders

Meta's Muse Spark Has No API Launch Date

By Alessandro Benigni

Microsoft's MAI bets enterprises will tune their own models — featured image for AI Insiders

Microsoft's MAI bets enterprises will tune their own models

By Alessandro Benigni

MiniMax M3 ships API-first with weights still pending — featured image for AI Insiders

MiniMax M3 ships API-first with weights still pending

By Alessandro Benigni

Alibaba's Qwen3.7-Plus aims to unify GUI and CLI in one agent loop — featured image for AI Insiders

Alibaba's Qwen3.7-Plus aims to unify GUI and CLI in one agent loop

By Alessandro Benigni

JetBrains ships Mellum 2, a 12B MoE model built for IDE-native coding — featured image for AI Insiders

JetBrains ships Mellum 2, a 12B MoE model built for IDE-native coding

By Alessandro Benigni

NVIDIA Ships Cosmos 3, an Open Foundation Model That Outputs Robot Actions — featured image for AI Insiders

NVIDIA Ships Cosmos 3, an Open Foundation Model That Outputs Robot Actions

By Alessandro Benigni

NVIDIA's Nemotron 3 Ultra is the strongest US open-weights model — featured image for AI Insiders

NVIDIA's Nemotron 3 Ultra is the strongest US open-weights model

By Alessandro Benigni

MiniMax M3 beats Opus 4.7 on BrowseComp as open-weights — featured image for AI Insiders

MiniMax M3 beats Opus 4.7 on BrowseComp as open-weights

By Alessandro Benigni

Nvidia's N1X laptop chip targets Intel's last stronghold — featured image for AI Insiders

Nvidia's N1X laptop chip targets Intel's last stronghold

By Alessandro Benigni

Prism ML ships 4B-parameter diffusion that runs on an iPhone — featured image for AI Insiders

Prism ML ships 4B-parameter diffusion that runs on an iPhone

By Alessandro Benigni

Anthropic releases Opus 4.8 with effort controls and cheaper fast mode — featured image for AI Insiders

Models / May 30

Anthropic releases Opus 4.8 with effort controls and cheaper fast mode

By Alessandro Benigni

Microsoft is reportedly building its own AI coding model — featured image for AI Insiders

Models / May 30

Microsoft is reportedly building its own AI coding model

By Alessandro Benigni

MiniMax teases M3 with sparse attention that runs 15.6x faster at long context — featured image for AI Insiders

Models / May 30

MiniMax teases M3 with sparse attention that runs 15.6x faster at long context

By Alessandro Benigni

NVIDIA's gamma-World adds independently controllable multi-agent rollouts — featured image for AI Insiders

Models / May 30

NVIDIA's gamma-World adds independently controllable multi-agent rollouts

By Alessandro Benigni

Sakana Labs claims a block-wise training method that bypasses backprop — featured image for AI Insiders

Models / May 30

Sakana Labs claims a block-wise training method that bypasses backprop

By Alessandro Benigni

Biohub ships open protein AI stack with 6.8 billion sequences — featured image for AI Insiders

Models / May 29

Biohub ships open protein AI stack with 6.8 billion sequences

By Alessandro Benigni

Callstack's Apex specializes a coding model for React Native — featured image for AI Insiders

Models / May 29

Callstack's Apex specializes a coding model for React Native

By Alessandro Benigni

ElevenLabs ships Music v2 with mid-track genre switching — featured image for AI Insiders

Models / May 29

ElevenLabs ships Music v2 with mid-track genre switching

By Alessandro Benigni

NVIDIA's LocateAnything decodes vision-language bounding boxes in parallel — featured image for AI Insiders

Models / May 29

NVIDIA's LocateAnything decodes vision-language bounding boxes in parallel

By Alessandro Benigni

A new repo tracks the shift to native multimodal models — featured image for AI Insiders

Models / May 28

A new repo tracks the shift to native multimodal models

By Alessandro Benigni

Claude Mythos finds OpenAI's Erdos disproof with a cleaner shorter proof — featured image for AI Insiders

Models / May 28

Claude Mythos finds OpenAI's Erdos disproof with a cleaner shorter proof

By Alessandro Benigni

DeepSWE introduces a coding benchmark that separates what SWE-Bench Pro cannot — featured image for AI Insiders

Models / May 28

DeepSWE introduces a coding benchmark that separates what SWE-Bench Pro cannot

By Alessandro Benigni

Microsoft AI's MAI-Image-2.5 enters Arena top 3 for text-to-image — featured image for AI Insiders

Models / May 28

Microsoft AI's MAI-Image-2.5 enters Arena top 3 for text-to-image

By Alessandro Benigni

AlphaProof Nexus solves nine open Erdos problems at a few hundred dollars each — featured image for AI Insiders

Models / May 27

AlphaProof Nexus solves nine open Erdos problems at a few hundred dollars each

By Alessandro Benigni

GPT 5.2 is the only model that can build a hard benchmark — featured image for AI Insiders

Models / May 27

GPT 5.2 is the only model that can build a hard benchmark

By Alessandro Benigni

GPT-5.6 reportedly launches in June with focus on reasoning and agentic flows — featured image for AI Insiders

Models / May 27

GPT-5.6 reportedly launches in June with focus on reasoning and agentic flows

By Alessandro Benigni

On-policy distillation closes the training gap between student and teacher — featured image for AI Insiders

Models / May 27

On-policy distillation closes the training gap between student and teacher

By Alessandro Benigni

Gemini 3.5 Flash (Low) generates 45% fewer tokens than Medium — featured image for AI Insiders

Models / May 26

Gemini 3.5 Flash (Low) generates 45% fewer tokens than Medium

By Alessandro Benigni

Mythos 1 surfaces in cloud security tools ahead of broader release — featured image for AI Insiders

Models / May 26

Mythos 1 surfaces in cloud security tools ahead of broader release

By Alessandro Benigni

Goodfire finds SAE features are map fragments, not atoms of meaning — featured image for AI Insiders

Models / May 22

Goodfire finds SAE features are map fragments, not atoms of meaning

By Alessandro Benigni

Microsoft in talks to supply Maia 200 chips to Anthropic — featured image for AI Insiders

Models / May 22

Microsoft in talks to supply Maia 200 chips to Anthropic

By Alessandro Benigni

Qwen3.7-Max arrives as a closed-weight agent foundation model — featured image for AI Insiders

Models / May 22

Qwen3.7-Max arrives as a closed-weight agent foundation model

By Alessandro Benigni

Alibaba unveils the Zhenwu M890 to chase AI agent workloads — featured image for AI Insiders

Models / May 21

Alibaba unveils the Zhenwu M890 to chase AI agent workloads

By Alessandro Benigni

ByteDance open-sources Lance, a 3B unified model for image and video — featured image for AI Insiders

Models / May 21

ByteDance open-sources Lance, a 3B unified model for image and video

By Alessandro Benigni

LiteFrame cuts video LLM inference latency 35% by shrinking the vision encoder — featured image for AI Insiders

Models / May 21

LiteFrame cuts video LLM inference latency 35% by shrinking the vision encoder

By Alessandro Benigni

Meta AI's WavFlow generates synchronized audio in raw waveform space — featured image for AI Insiders

Models / May 21

Meta AI's WavFlow generates synchronized audio in raw waveform space

By Alessandro Benigni

No-filter pretraining outperforms curated data at scale — featured image for AI Insiders

Models / May 21

No-filter pretraining outperforms curated data at scale

By Alessandro Benigni

OpenAI model disproves Erdos geometry conjecture from 1946 — featured image for AI Insiders

Models / May 21

OpenAI model disproves Erdos geometry conjecture from 1946

By Alessandro Benigni

Stability AI ships Stable Audio 3.0 with open weights and six-minute tracks — featured image for AI Insiders

Models / May 21

Stability AI ships Stable Audio 3.0 with open weights and six-minute tracks

By Alessandro Benigni

AllenAI ships OlmoEarth v1.1 with 3x compute reduction — featured image for AI Insiders

Models / May 20

AllenAI ships OlmoEarth v1.1 with 3x compute reduction

By Alessandro Benigni

Google launches Gemini 3.5 Flash with 4x speed edge over rival frontier models — featured image for AI Insiders

Models / May 20

Google launches Gemini 3.5 Flash with 4x speed edge over rival frontier models

By Alessandro Benigni

NVIDIA open-sources LongLive 2.0 for real-time interactive video — featured image for AI Insiders

Models / May 20

NVIDIA open-sources LongLive 2.0 for real-time interactive video

By Alessandro Benigni

A 1B model that costs $1,500 to train from scratch — featured image for AI Insiders

Models / May 19

A 1B model that costs $1,500 to train from scratch

By Alessandro Benigni

Alibaba's Qwen3.7 enters Arena rankings at positions 13 and 16 — featured image for AI Insiders

Models / May 19

Alibaba's Qwen3.7 enters Arena rankings at positions 13 and 16

By Alessandro Benigni

NVIDIA hand-delivers Vera CPUs to Anthropic, OpenAI, and Oracle — featured image for AI Insiders

Models / May 19

NVIDIA hand-delivers Vera CPUs to Anthropic, OpenAI, and Oracle

By Alessandro Benigni

LLMs Switch Modes Mid-Training in Ways Optimizers Cannot Fix — featured image for AI Insiders

Models / May 19

LLMs Switch Modes Mid-Training in Ways Optimizers Cannot Fix

By Alessandro Benigni

Censorship in Qwen 3.5 Is a Removable Circuit, Not a Knowledge Gap — featured image for AI Insiders

Models / May 19

Censorship in Qwen 3.5 Is a Removable Circuit, Not a Knowledge Gap

By Alessandro Benigni

Featured image for: Anthropic limits Claude Mythos to a closed security consortium

Models / May 19

Anthropic limits Claude Mythos to a closed security consortium

By Alessandro Benigni