Trend canvas ai

everyday tech, extraordinary creativity.
simple hindi stories, daily.
made with ai, checked with care.

[Trend canvas ai @Trendcanvasai]

NVIDIA ke Hugging Face page ke mutabik Nemotron-3.5-Bolt-E4B Gemma-4-E4B se post-train kiya chhota model hai (8B total, 4B effective), local voice aur device assistants ke liye. 64K context, license OpenMDW-1.1.

https://t.co/48Q3h21wBb

#nemotron #ai #opensource

[Trend canvas ai @Trendcanvasai]

KRAFTON AI ke Hugging Face page ke mutabik Raon-1 mini preview ek 30B-A4B MoE model hai (28.8B total, 4.3B active), Korean aur agentic tasks par focus. Apache 2.0, 262K context. Preview release hai, specs badal sakte hain.

https://t.co/4stpqlfhC1

#raon #opensource #ai

[Trend canvas ai @Trendcanvasai]

Shanghai AI Lab ke Hugging Face page ke mutabik Atria Dawn Preview ek agentic model hai (744B MoE GLM-5.2 par bana), research aur engineering tasks ke liye. 256K context, MIT license.

https://t.co/jurjyLRx6r

#atriadawn #opensource #ai

[Trend canvas ai @Trendcanvasai]

Apache Maka (Incubating) ke README ke mutabik yeh ek agent workspace hai jo har model message, tool call aur permission decision ko append-only log mein rakhta hai. Aap apna model lao: cloud API ya local.

https://t.co/jSSWYEFANn

#apache #aiagents #opensource

[Trend canvas ai @Trendcanvasai]

MemTensor ke GitHub ke mutabik MemT-1.0 memory-native model hai: memory model ke compute ka hissa hai. Unka dava: 128K history par state size fixed, KV cache se ~66x kam storage (company-reported).

https://t.co/ufnIs681Br

#memt #llm #ai

[Trend canvas ai @Trendcanvasai]

DeepSeek ke README ke mutabik DeepInfer unka production inference framework hai, ab open-source (Apache-2.0). Abhi offline inference ke saath DeepSeek-V4.1-Flash supported hai.

https://t.co/LlgHVwEyQV

#deepseek #inference #ai

[Trend canvas ai @Trendcanvasai]

Epic Games ke README ke mutabik Lore ek open-source (MIT) version control system hai, jo code ke saath bade binary assets, jaise games, ke liye bana hai. Abhi pre-1.0 hai, yaani interfaces badal sakte hain.

https://t.co/3y1EtgpGFT

#epicgames #opensource #git

[Trend canvas ai @Trendcanvasai]

Linux Foundation ke mutabik 22 Oct ko Agentic AI Foundation mein 4 naye projects aaye: Agent Skills (Anthropic), Open Responses (OpenAI), OpenEnv (Hugging Face), OpenSandbox (Alibaba). Ab AAIF mein 10 projects hain.

https://t.co/r2ahEHZS3a

#aaif #opensource #ai

[Trend canvas ai @Trendcanvasai]

Google ke README ke mutabik Elephant-Goldfish Model AI coding agents ke liye open-source spec-driven plugin hai: "Elephant" session spec likhta hai, zero-context "Goldfish" subagents code se pehle use stress-test karte hain.

https://t.co/S8qgsBq4HG

#google #aicoding

[Trend canvas ai @Trendcanvasai]

Nanbeige ke model card ke mutabik Nanbeige4.3-33B open (Apache-2.0) agentic coding model hai, jo chhote agent models ko distill karne ke liye "teacher" ki tarah banaya gaya. ~30B class mein best hone ka dava company ka apna hai.

https://t.co/vdSYsShFUD

#nanbeige #opensource #ai

[Trend canvas ai @Trendcanvasai]

Tencent ki project page ke mutabik Hy-Nomas 8B-parameter interactive world model hai: text aur camera control se 720p tak, maximum 60 sec ka video banata hai. Weights Apache-2.0 (third-party parts ke alag license).

https://t.co/m1Rh6H5qPO

#tencent #worldmodel #ai

[Trend canvas ai @Trendcanvasai]

Hugging Face ke model card ke mutabik Carbon-X 14.8B ka open (Apache 2.0) model hai jo chat mein raw DNA padhkar English mein jawab deta hai. Research starting point hai: clinically validated nahi, medical decisions ke liye nahi.

https://t.co/nrFq82afhk

#huggingface #ai

[Trend canvas ai @Trendcanvasai]

PyTorch blog ke mutabik TorchTPU Google aur Meta ka open-source collaboration hai: TPU ke liye native PyTorch backend, runtime mein JAX dependency nahi. Goal: GPU se TPU par port karna bas device badalne jitna aasaan.

https://t.co/WfWX5ZIqbu

#pytorch #tpu #ai

[Trend canvas ai @Trendcanvasai]

AWS ka Agent Toolkit (Apache-2.0): official MCP servers, skills aur plugins, jo Claude Code, Codex, Cursor, Devin aur Kiro mein chalte hain. aws-core plugin CDK/CloudFormation, serverless, billing jaisi cheezein cover karta hai.

https://t.co/yZB8sGBE5y

#aws #mcp #ai

[Trend canvas ai @Trendcanvasai]

Moonshot ne Kimi K3.1 Pro mein cyber-offensive ability jaan-boojh kar kam ki: training data se cyber data hataya. Unka internal test: 81.0% vs 90.5% bina curbs ke (company-reported, unverified). Weights 4 Nov tak.

https://t.co/DfLGBgydS7

#kimi #aisafety #ai

[Trend canvas ai @Trendcanvasai]

Moonshot ne Kimi K3.1 Pro mein cyber-offensive ability jaan-boojh kar kam ki: training data se cyber data hataya. Unka internal test: 81.0% vs 90.5% bina curbs ke (company-reported, unverified). Weights 4 Nov tak.

https://t.co/DfLGBgxG2z

#kimi #aisafety #ai

[Trend canvas ai @Trendcanvasai]

Cerebras ke blog ke mutabik 3T Kimi K3 ko 59 wafers par baantna padta hai: 35 wafers linear-attention layers ke liye, 24 dense ke liye (4-bit weights, 256k request maan kar; illustrative). 1000-2000+ tok/s unka apna claim hai.

https://t.co/LfpQRAwbgG

#cerebras #llm #ai

[Trend canvas ai @Trendcanvasai]

@evgheni_D Anthropic's post gives no selection rule beyond using existing OSS-Fuzz fuzz tests: hundreds attempted, 239 proved. Outside reach, per them: bugs ASan misses, unverified entry points and dead code, caller misuse, and flaws in the Rocq/RISC-V/Linux models.

[Trend canvas ai @Trendcanvasai]

Anthropic ke mutabik, Claude ne kuch hafton mein 100+ open-source libraries ki security properties prove ki - pehle formal verification sirf chhote codebases mein possible tha. Hamara take: AI ab sirf code likh nahi raha, use saabit bhi kar raha hai.

#ai #security #anthropic

[Anthropic @AnthropicAI]

New on our Frontier Red Team blog: using Claude to formally verify software at large scale.

Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time).

With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks.

Read our research in full here: https://t.co/zUyx8CjBEk

[Trend canvas ai @Trendcanvasai]

Microsoft ka Agensh open-source: bina central orchestrator 1,024 coding agents. Pandoc pass rate: 1 agent 33.89%, 1,024 agents 55.06% (GPT-5.6-sol). Par ek GPT-6.1-sol agent 56.19% tha. Hamara take: agents se pehle model dekho.

https://t.co/eu4FSECbu4

#agents #opensource #ai

[Trend canvas ai @Trendcanvasai]

AI agents robots ko chalna sikhane mein bhi haath bata rahe hain. NVIDIA ke mutabik agents ne MicroDuck ka obstacle course plan kiya; Isaac Lab mein 4,096 parallel sims mein policy train hui. Abhi sim mein; asli robot agla kadam.

https://t.co/QtMEe3fVYH

#nvidia #robotics #ai