Hina Dixit

BlackBird, you beauty! Comparison side to side: BlackBird vs Perplexity, where BlackBird is completely local and uses much smaller model. Still in Beta but you can request access here: lnkd.in/gXBTtWnX Follow @DecomputeAI for more updates.

Only top 1% people in AI would know what’s happening here. Comment if you know. #AI #SLMs #Model #LLM #ArtificialInteligence

Read our blog here to learn more about the release: shorturl.at/5XOMc

Small models don’t just need more training. They need better reasoning architecture. We introduce novel model architecture: Nebula-S-SVMS1-4B: HMMT: 66.7% GSM8K: 90% MMLU-Pro: 79.7% GPQA: 70.5% Reasoning-native, on-device AI is coming. #OnDeviceAI #Reasoning #EdgeAI #SLM

Introducing BlackBird Enterprise — a structurally private AI platform that runs where your data already lives: on-device, on-prem, or in your cloud. Built for orgs that refuse to compromise on: • Data residency, compliance & guardrails • Structural privacy by design • On-device/on-prem RAG, document AI, copilots • Distributed inference & training (DGX, multi-tenant) • Works with files, folders, KBs, Excel, SharePoint, GDrive, Confluence, Notion • Voice → transcription, summaries, Q&A • Powered by our Nebula-S models beating Qwen & DeepSeek on MMLU/ARC/GSM8K If you’re navigating AI + privacy + cost, this is for you 👇 lnkd.in/gpQViNkn #AI #EnterpriseAI #PrivacyByDesign #OnDeviceAI #GenAI

Task: I need to schedule three more meetings tomorrow. 🤓 Problem: I only have 24 hours a day. 🧐 Founder life: back-to-back customer calls, no time for lunch, and 87 unread emails. Living the dream. 😎

@baelentina_ 🔥

super serious Everything Shower routine dropping soon

What if AI could learn a new task just by watching? @DecomputeAI’s Kestrel VLM learns complex visual tasks from only a few demonstrations, bridging perception, reasoning, and action in one model. Built for the edge. Designed for the real world. Thank you to my wonderful team for putting this demo together.

Kestrel VLM is holding its own against Apple’s FastVLM. 📊 RealWorldQA accuracy: •Kestrel-1B → 62.3% •Kestrel-2B → 63.8% •FastVLM-0.5B → 56.1% •FastVLM-1.5B → 61.2% •FastVLM-7B → 67.2% Smaller models, competitive accuracy. Fusion-space pays off. @DecomputeAI 🙌🏻

Most businesses, teams and users need to teach a model a brand‑new workflow the moment it’s deployed, often with only 5-10 examples on‑device examples but today they’re stuck between options that don’t actually learn or don’t fit at the edge. In‑context “few‑shot” prompting never updates weights, so the model forgets as soon as the prompt disappears. One‑off fine‑tunes and RL‑style alignment push data to the cloud and balloon memory by keeping long reward graphs and activations alive, routinely overwhelming laptop‑class GPUs and clashing with privacy constraints. What’s missing is an on‑device path that turns a handful of demonstrations into persistent skill and keeps improving from everyday use, without exceeding a minute memory envelope.  Follow us @DecomputeAI if you are interested in this problem space!

@jxmnop Try AtlasTune by @DecomputeAI. We reduce trainable parameters by 2000X compared to LoRA.

@deedydas Congratulations! Well deserved!

Big personal news! After a remarkable 18mos at Menlo Ventures, I'm excited to announce that perhaps against better judgement, they have decided to make me a Partner at the firm. Wild that just 6yrs ago, I'd be up at 2am debugging Google Search. A little bit about my journey:

@rohanpaul_ai Couldn’t agree more. We @DecomputeAI saw this opportunity last year, and have made immense progress including launch of Kestrel SLMs which self-improve at the edge.

This is probably one of THE most important paper of the last few months. Small language models are sufficiently powerful, operationally suitable, and economical Agentic tasks. - Phi-2 matches 30 billion models running 15x faster. - Serving a 7 billion parameter small language model is 10–30x cheaper than larger models. - Agentic applications use only limited language model capabilities, fitting well with specialized small models. - Heterogeneous systems use efficient small models routinely, invoking large models sparingly for general tasks. - A conversion process is recommended that involves logging agent interactions, clustering tasks, selecting small models, and fine-tuning them on task-specific data. SLM fine-tuning aligns behavior precisely for structured agent interactions like tool calls. Heterogeneous systems blend SLM efficiency for routine tasks with LLM power for complex steps. On-device SLM inference delivers low latency and enhanced data privacy for agent users. --- Paper - arxiv. org/abs/2506.02153 Paper Title: "Small Language Models are the Future of Agentic AI"

🔥 New Release on high demand: Now train Open AI’s GPT-OSS 20B for personalized agents locally on Mac with BlackBird on demand: ⚡ On-device training 🔒 Full privacy 🛠️ No cloud, no compromise decompute.run/releases (We received several messages from everyone requesting this release, and it seems MLX was completely broken and did not support fine-tuning, so we fixed it ourselves.) #BlackBird #Decompute #OpenAI #oss

@MohapatraHemant My OCD.

What's stopping you from living like this...

2/ Why it matters: OCR in the wild, scene reasoning for real-world QA, and DocVQA in one model reduces brittle pipelines. Give it a try here: decompute.run #VisionAI #AI

1/ Kestrel VLM 1.5B — 841/1000 on OCR, 63.5% on Real-World QA and 89% on DocVQA eval!! One compact model, three hard tasks. #VLM #OCR #DocVQA #Multimodal #SLM

More improved with a score of 841/1000. Now hosted: decompute.run

We recently shipped Kestrel VLM - a self-improving, optimized Vision-Language model for macOS, Windows, Android and iOS, available in three sizes (650M, 1B and 1.5B). Achieves ~817/1000 for OCRB, rivaling 32B-param giants, yet runs entirely on-device. Perfect for document understanding, screenshot parsing & vision-driven agents, and it keeps getting smarter after deployment with a memory footprint of under 5GB. Very proud of the team for the breakthrough performance! 👏🏻👏🏻👏🏻

🚨 Calling all builders in SF! $50K up for grabs! 🚨 We’re thrilled to sponsor the A2A Agents Hackathon in SF! 💥 Build on-device AI agents with the Blackbird SDK 🧠 Multimodal, fast, secure, insanely powerful 🏆 Judged by Decompute’s own @hinadixit 🎯 Sign up: lu.ma/aoc5opn4 #AIAgents #Hackathon #OnDeviceAI #Decompute #BlackbirdSDK

📢 Exciting news! The BlackBird-SDK is now open source! 🎉 - Local AI inference | - Real-time streaming | - Secure file processing 👉 Contribute: github.com/Decompute-AI/B… #OpenSource #AI #MachineLearning