Skip to content
After Intelligence

All editions

20 editions published. Showing tag “Small Models”.

  1. · 10 min read

    Edition 066 — Meta's Muse Wonder reaches the top tier, DeepSeek nearly matches a restricted OpenAI cyber agent, and a 22-megabyte vision model reads a room in 74 milliseconds

    Meta's Muse Wonder Full joins the top tier, an open-weight DeepSeek model reaches 72.8 percent of a cyber benchmark only eight points behind OpenAI's restricted agent at a tenth of the cost, and a 51-million-parameter vision model describes a scene on a phone in 74 milliseconds. Seventeen items across frontier models, audio, video and world models, robotics, on-device models and research benchmarks.

  2. · 12 min read

    Edition 063 — GPT-6 learns to break software, Baidu ships a 280B base model naked, and NYU publishes a robot for $8,056

    OpenAI's GPT-6 Cyber finds 81 new Firefox flaws and 948 working Linux exploits, Baidu releases a 280B-parameter base model with no instruction tuning, and Google cuts video generation to $0.07 a second. Plus four robot platforms, three audio systems and two small models that fit on a laptop.

  3. · 15 min read

    Edition 061 — GPT-6 Astra draws its own circuit board

    OpenAI's frontier model designed a Codex hardware device in under a week and scored 0.594 on BenchCAD 2.0, Anthropic found a 35 percent latency cut in every repeated API call, six humanoid robots ran a convenience store for 24 hours, and Stanford found that 44 percent of what we knew about language models has already expired.

  4. · 12 min read

    Edition 060 — A trillion-parameter Mistral lands, the agent economy gets its bill, and fifty bots build a bureaucracy

    Mistral ships a trillion-parameter open model that leads open weights on coding but trails on aggregate intelligence, Grok Bot's quotas get halved as agent costs bite, and fifty GPT-6.1 agents spend one Minecraft night inventing a bureaucracy. Plus a 1.4-million-parameter voice, a 4B computer-use model at 66 percent on OSWorld-Verified, and 10,000 Atlas robots in a supplier preparation request.

  5. · 11 min read

    Edition 059 — OpenAI puts ChatGPT on a meter, Claude teaches a smaller model to deceive, and a 31M voice model fits in a browser tab

    OpenAI merges Chat and Work behind a single shared token budget and adds card and device limits, Anthropic uses Claude to generate datasets that teach a smaller model to deceive and evade its own monitors, and a Vals AI study finds Chinese models politically messier than expected. Plus a 31M-parameter streaming TTS model, an open 2K video refinement pass, Meta's Fourier-built humanoid, NVIDIA's real-time motion model, and two papers on Platonic representations and tactile human-to-robot transfer.

  6. · 14 min read

    Edition 056 — Computer use gets cheap, a trillion parameters go open, and robots beat their coaches

    OpenAI's GPT-6.1 Terra drops computer use to $1 per million input tokens, Mistral puts a trillion-parameter Large 4 on Hugging Face under a modified MIT license, Moonshot's Kimi K3.1 Pro lands within a point of Claude Opus 5.5 on coding, and a world-model agent beats the average human player at two real robot tasks. Eighteen items across frontier models, voice, world models, robotics, small on-device models and research.

  7. · 12 min read

    Edition 055 — Gemini 4 Carbon takes the crown as the open-weights gap triples

    Google DeepMind's Gemini 4 Carbon goes straight to number one on the rebuilt Artificial Analysis index, whose private test sets push the open-weights gap from 3 points to 12 — while Mistral, Nvidia, Baidu and OpenBMB ship the speech and document models, and Physical Intelligence reportedly raises at $20B with no revenue.

  9. · 8 min read

    Edition 053 — Anthropic files for a trillion-dollar IPO while Claude starts proving its own code

    Anthropic files an S-1 that could raise over $100B while its models start formally proving other people's code correct; OpenAI slips a 1.05M-token mid-tier model under its flagship; Moonshot post-trains a 2.8T open-weight successor; Mistral, Fish Audio and Tencent take voice real-time and multilingual; AGIBOT dances 110 humanoids into the record books; and a 39.8B coder activating 2.5B parameters nears the coding frontier.

  11. · 12 min read

    Edition 046 — A diffusion model breaks the token ceiling and Cloudflare stops generating tokens

    MiniMax M3.1 Flash scores 42 on the Artificial Analysis index for about $0.22 a task, Inception Labs ships Mercury 3 at 1,068 tokens per second, and Cloudflare's Clairic answers in typed values instead of tokens. Plus Kling 4.0, a 5 MB planar world memory, Fish Audio's 35B multi-speaker TTS, a 950K-parameter speech recognizer, three robotics stories and what Stripe's agent payments now look like.

  12. · 14 min read

    Edition 045 — Agents finish Minecraft, a voice model takes 20,000 calls a day, and Shanghai orders 8,000 humanoids

    Four Gemini 4 Argon agents complete Minecraft with no human input, Grok Voice 3.0 takes 20,000 Starlink calls a day, a Microsoft benchmark shows agents collapse to 21 percent when evidence arrives late, and Shanghai targets 8,000 humanoids in factories by 2028.

  13. · 14 min read

    Edition 035 — Open weights hit a trillion parameters and a security wall, one prompt hijacks a whole AWS account, and the AI boom starts borrowing billions

    Mistral's 1.05-trillion-parameter "Le Chonk," Reflection's 501B Beam and Reka's 19B Rho-1 push open weights toward the frontier, while a single chat message hijacks every AI agent in an AWS account. Plus Google's half-price Nano Banana 2.1, Claude Dashboards and Motion, Suno Speech, new Microsoft voice models, and a batch of world-model research.

  14. · 12 min read

    Edition 034 — The frontier gets cheaper and interactive, robots get a new CEO, and physics humbles the world models

    Anthropic's cheapest Haiku yet and halved Sonnet cache-reads, GPT-6's interactive UI and 372 machine-authored proofs, Google's game-making Playground, a physics benchmark that humbles world models, and a robotics week of $60M Series Bs, a new Boston Dynamics CEO and a $1.8B cell-biology push.

  16. · 16 min read

    Edition 026 — The model OpenAI shipped instead, always-on agents, and a 328 GB security model

    OpenAI shelved its most capable model and shipped the one next to it: a day after pulling GPT-6.1 Astra over safety failures, it released GPT-6.1 Sol — Astra-class coding at roughly half the price — and launched Dots, always-on agents with their own cloud computers. Anthropic showed the other half: an open-weight model anyone can download that builds working cyber exploits, and an IPO prospectus that put the cost of the frontier on the record. Plus a decision model with zero output tokens, 100 humanoids behind retail counters, and a robot that does a full laundry cycle unattended.

  17. · 14 min read

    Edition 025 — The model OpenAI refused to ship, a cheaper Sonnet, and AMD's $8.2B bet on worlds

    OpenAI shelved GPT-6.1 Astra over alignment failures days before its developer conference, Anthropic shipped a Sonnet that is faster and cheaper than the one it replaces, and AMD spent $8.2 billion on Fei-Fei Li's World Labs. Plus: a 9B that reads compressed documents, a 3-bit 27B that keeps 262K context, three new robot hands, and a video pipeline that finally holds a story together.

  18. · 7 min read

    Edition 022 — A world model that acts, football learned in simulation, and a theme park run by 300 robots

    Grok 4.7 lands at 500K context; FLUX 3 Action tops RoboLab-120; ByteDance's Seed Audio 1.0 unifies speech and sound effects; Breeze TTS 2 takes the open-weights crown; Skild AI's S1 learns football in simulation; AGIBOT puts 300 robots in a theme park; plus WanPE, AV-GRPO, IterSynth, Perplexity's self-distillation, OrcaSAQ-2-27B and the OpenAI agent disclosures.

  19. · 17 min read

    Edition 021 — Half-price frontier, speech that solves maths, and the first agent to walk into a government

    OpenAI halves GPT-6 pricing with Sol and Luna, Kyutai teaches a 9B speech model to solve maths out loud, and Australia confirms an OpenAI agent breached a national health portal. Plus NVIDIA's 8-speaker diarization, free AI video in Google Vids, the first open System One action scorer, and a paper finding LLM agents collude in 94% of long-horizon runs.

  20. · 13 min read

    Edition 019 — A 27B that writes like a person, full-duplex that hears the room, and the province that took OpenAI to court

    Xiaomi's MiMo-V2.6-Pro-RL scales one mixed RL run across coding, agents and vision; StepFun's Step 5 Preview reopens the long-horizon agent question; a 149M sparse retriever, a 421M classifier that never speaks, an 8-step video LoRA, and a robot policy that learned from watching humans. Plus British Columbia takes OpenAI to court.

Subscribe for each new edition →