Building https://teapaw.org/t.co/6spQ572OL1 + https://teapaw.org/t.co/rJw10GFmtJ + https://teapaw.org/t.co/uxWWE30GPC & ++ Founding member of @CryptoKitties @flow_blockchain @klktnofficialJP
Vancouver
Joined March 2010
- Tweets16.8K
- Following1K
- Followers1.1K
- Likes15.8K
Fab retweeted
Introducing π»π΅πΌπΉ: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> π»π΅πΌπΉ-π·.π½π±: text, 256k context window.
> π»π΅πΌπΉ-π
π»-πΈπ±: vision-language model with video-input support for spatial understanding
> π»π΅πΌπΉ-π°ππππ-πΈπ±: tool calls, voice chat, reasoning in 18 languages.
> π»π΅πΌπΉ-πΎπππ-πΈπ±: an omni model combines vision, audio, and language.
> ππ·-πΈπ± and ππ·-πΎπππ-πΈπ±: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> π»π΅πΌπΉ-π
π»-πΈπ± is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> π»π΅πΌπΉ-π·.π½π± is #1 on the Pareto frontier of device intelligence πππππππ.ππππππ.ππ
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ π§΅
Fab retweeted
China's 2008 is happening now.
Property investment in China dropped -21% YoY in the first 9 months of 2026, its largest decline for this period of the year on record.
This marks a deterioration from a -20% YoY decline in the first 8 months of 2026.
The latest drop is also significantly worse than the -16% decline recorded during the 2020 pandemic.
Property investment in China has now contracted every month since 2022, its longest declining streak on record.
For perspective, during the 2008 Financial Crisis, this figure never fell below +1% YoY growth.
China's property sector is experiencing a historic downturn.
We're entering the Ready Player 1 era - very quickly!
So many things don't even need to wait models to generate anymore, pick from one of the many free 3D libraries.
We just open-sourced 1,300+ 3D models.
All 39 GB are free to download on Hugging Face.
huggingface.co/datasets/scenβ¦
> made with Hitem3D, Meshy, Rodin, Tripo, Claude Opus 5.5 3D and GPT-6 Astra 3D, which all run on Scenario
> includes the input images, prompts and settings that made them
> GLBs up to 8K textures, native Blender files from Opus and Astra, CC BY 4.0
When you get to feel that God is truly the entirety of creation, the universe, the stars, you and me and everything really - it's far less glamorous than it once sounded.
It's one overarching entity playing all the roles in a cosmic theater while also being the audience at the same time.
Don't take me wrong, it's incredibly mind-boggling still. It's just that in a single tangible dimension, it's comprehensible and relatable.
If truly single shot here, looking at details (models, motion, shadows) - Kimi 3.1 did WAY better.
KIMI K3.1 PRO IS LITERALLY A CHEAPER AND FASTER VERSION OF OPUS 5.5
i asked both models to build the exact same animated voxel 3D pagoda island
- intricate pagoda architecture, bridges and floating islands
- detailed trees and tiny environmental elements
- Kimi: 15 mins vs Opus: 32 mins
- Kimi: $5.71 vs Opus: $17.13
roughly 2x faster and 3x cheaper for a comparable result
who wouldβve thought chinese AI models would be challenging SF labs this soon π
Makes me think.. is the future exclusively MoE models? π€
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
πΉ 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
πΉ Full-project coding: games in Unreal, Unity & Godot.
πΉ Excels at parametric 3D modeling in Blender & CAD.
πΉ More reliable at finance work: Accurate modeling with hallucination down by 65%.
πΉ Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2β3Γ faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
π API: platform.kimi.ai
π Tech blog: kimi.ai/blog/kimi-k3-1-pro
PI + Deepseek is magic, we do have Opus at home now.
I repeated DeepSeek V4.1 Flash 1017 benchmark 89 times across 9 harnesses to understand which works best.
High effort, 5 attempts each:
Pi harness: 40.4
Grok Build CLI: 40.4
DeekSeek: 39.2
OpenCode 2.0: 38.6
OpenCode: 36.6
Claude Code: 36.4
Codex: 33.2
Hermes Agent: 32
Kilo Code: 31.8
Codex kept failing several times exceeding the allowed time per repo (3 hours) and was judged where it finished.
Based on what I tested, Pi, Grok Build CLI, and OpenCode 2.0 look like the strongest universal harnesses. This doesn't necessarily generalize.
A detailed table with every run in the responseπ§΅
Fab retweeted
Pi harness shows a similar trend while being cheaper:
DeepSeek harness: 42.5 (mean of 7)
Pi harness: 42.2 (mean of 5)
Claude Code harness: 39.1 (mean of 7)
All data on the benchmark website. I'm testing high effort + Pi next.
Fab retweeted
we architected the model to read from inputs in a single forward pass. this makes it incredibly cheap and fast when compared to LLMs, since there is no token generation involved.
in our live demo below, you can see how using a cascaded pipeline with clairic + clef + LLM is 3.5x faster and 77% cheaper than just using an LLM.
clairic-pipeline-demo.workerβ¦
We need more behavioural benchmarks!
AI λͺ¨λΈλ€ λμ¬ μΈ‘μ ν λ²€μΉλ§ν¬ λμλλ° κ²°κ³Ό κ°μκΉγ
γ
GPT-4oλ κ·Όκ±° μλ λ°λ°μλ 94.6%λ μ견μ λ°κΏμ κ·Έλ₯ κΈμ λ΄ μμ€μ΄κ³ γ
γ
λ°λλ‘ Opus 5.5λ νλΉν λ°λ°μ λ°μλ€μΈ λΉμ¨μ΄ κ³ μ 32.2%μ.
μλ μ§μ§ μκΈ°κ° νλ €λ λ§μ μ λ£λ 건κ°γ
γ
γ
νλ¦° κ±° μ§μ ν΄ μ€λ "μλλ°? λ΄κ° λ§λλ°?" μμ νλ κ³ μ§μ΄ 보ν΅μ΄ μλγ
γ
λ°λ©΄ GPT-6 Astraλ νλΉν λ°λ°μ 88.3% λ°μλ€μ.
λ¬Όλ‘ νΉμ λ¬Έμ νκ° νκ²½μμ λμ¨ κ²°κ³Όμ§λ§, μ첨 μ νλ κ±°λ λ§μ΄ ν΅νλ 건 λ³κ°μΈ λ―ν¨ γ
γ
λ
Όλ¬Έ: arxiv.org/abs/2610.15634
Wow! This is a ridiculously accurate guess about what @typesafeai 's JEV from a really vague explanation - good job @claudeai Fable 1.0 - the OG!
What a cool time to be alive.
I remember so clearly 12y ago i was making a game on the `new` HTML5 <canvas> and had all sorts of concerns around capabilities it could support at the time.
All along, Apple's Safari actively undermining compatibility π
This repo was part of the effort -
github.com/flockonus/javascrβ¦
This was the game:
github.com/flockonus/gartsb
