Last time we shared a step-by-step installation guide for setting up the K2-Think model locally.

This time, we’re taking it further → we just published a brand-new AI Agent Building Guide powered by K2-Think, a 32B reasoning model created by UAE’s MBZUAI (Mohamed bin Zayed University of Artificial Intelligence) and G42.

K2-Think is designed for tough reasoning tasks in math, code, and science. It ranks high on benchmarks like AIME, HMMT, and LiveCodeBench, making it a powerful open-weights alternative for advanced problem solving.

What’s inside this new guide:
✅ Building a Math Dueler Agent with two proposers + one referee.
✅ Setting up environment & dependencies.
✅ Writing modular agent scripts.
✅ Integrating Sympy for math verification.
✅ Wrapping everything in a clean Gradio interface.
✅ Launching the app locally on your GPU VM.

Already covered setup & installation? Perfect. Jump straight into this agent guide.

Link: https://nodeshift.cloud/blog/building-a-math-dueler-agent-with-k2-think-step-by-step-guide

Also worth noting → K2-Think is available on NodeShift Sovereign Cloud and NodeShift AI, making it easy to run on trusted infrastructure.

Create complete, creative, intelligent visuals with just a simple text-prompt with Tencent's latest HunyuanImage 3.0.

With an 80B Mixture-of-Experts engine and a unified autoregressive framework, it delivers photorealistic, fine-grained images that don’t just follow the prompt, but also reason with them. Sparse prompt? No problem. This model fills in the gaps with world knowledge to produce visuals that feel intentional, accurate, and breathtakingly real.

With NodeShift Cloud’s one-stop GPU platform, you can set up and run HunyuanImage 3.0 effortlessly, skipping the hardware headaches while scaling creativity on demand.

🔗 Checkout our step-by-step guide: https://nodeshift.cloud/blog/how-to-install-and-run-hunyuanimage-3-0?utm_source=telegram&utm_medium=social&utm_campaign=hunyuanimage3

Tencent just released something crazy — and we built a full guide around it!

Introducing Hunyuan3D-Omni — Tencent’s newest unified image-to-3D generation framework.

This isn't your average text-to-3D tool. Omni lets you control the generation process with:
✅ Point Clouds
✅ Voxels
✅ 3D Bounding Boxes
✅ Skeletal Poses

All through a single control encoder, with options like EMA for smoother results and FlashVDM for faster inference. Runs perfectly with just 10–12 GB VRAM.

In this step-by-step guide, we’ve covered:
✅ GPU requirements
✅ How to set it up on a NodeShift GPU VM
✅ Exact commands to run point, voxel, bbox, and pose-controlled generation
✅ Output formats, inference tips, and more!

Whether you're in gaming, research, or 3D design — this model is worth a spin.

Check out the full tutorial here: https://nodeshift.cloud/blog/how-to-install-run-hunyuan3d-omni-locally

GLM 4.6, the latest release from Zai Org is an AI model that reasons, codes, and acts with unmatched power against some well known names like DeepSeek V3.1 Terminus and Claude Sonnet 4 .

Built on the next-gen GLM-4.6 foundation, it brings:
- 200K token context window – tackle complex tasks like never before
- Superior coding & agent performance – from Claude Code to Roo Code
- Advanced reasoning & tool use – stronger, smarter, more capable agents
- Refined human-aligned writing – natural style and role-playing scenarios

Our latest publish walks you through how to install & run GLM-4.6 locally or on GPU-accelerated environments with copy-paste ready steps.

🔗 Read here: https://nodeshift.cloud/blog/how-to-install-and-run-glm-4-6?utm_source=telegram&utm_medium=social&utm_campaign=glm46_launch

Kwaipilot just released KAT-Dev-32B — a powerful open-source coding assistant

KAT-Dev-32B (Kwaipilot/KAT-Dev) is a 32.8B-parameter coding assistant based on Qwen3-32B, purpose-tuned for software engineering tasks.

It’s trained in three stages — mid-training (core skills), SFT + RFT (teacher trajectories), and large-scale agentic RL (prefix caching + trajectory pruning + scalable infra).

On SWE-Bench Verified, KAT-Dev-32B achieves comparable performance with 62.4% resolved and ranks 5th among all open-source models with different scales.

We just published a step-by-step guide on how to set up and run KAT-Dev-32B on a GPU-powered NodeShift VM.

In this guide, we cover:
✅ GPU configuration requirements (single-GPU, multi-GPU, quantized setups)
✅ Step-by-step process to launch a NodeShift GPU VM
✅ Setting up JupyterLab with CUDA & PyTorch ready-to-go
✅ Installing libraries (Torch, Transformers, Accelerate, Einops)
✅ Running KAT-Dev interactively inside a notebook
✅ Generating your first response with the model

Check out the full guide here: https://nodeshift.cloud/blog/how-to-install-run-kat-dev-locally

MinerU2.5-2509-1.2B — A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing

MinerU2.5 is a compact 1.2B VLM with a smart two-stage, coarse-to-fine pipeline (global layout → native-res crops) that delivers state-of-the-art doc parsing with low compute. On OmniDocBench it tops the charts—overall 90.67, leading on Text (95.34), Formula (88.46), Table (88.22), and Reading Order (96.62)—outperforming many larger OCR/VLM systems.

What’s inside our new guide
✅ Setup end-to-end on a GPU VM (we demo with NodeShift, works anywhere)
✅ Two paths: Transformers (simple) & vLLM (fast + scalable, async engine ready)
✅ Copy-paste scripts to run two_step_extract() on your pages
✅ VRAM sizing & perf tips (quantization, token budgets, image sizing)
✅ Outputs you can use: structured blocks → Markdown, tables, formulas

Read the guide here: https://nodeshift.cloud/blog/how-to-install-run-mineru2-5-2509-1-2b-locally

Struggling to get AI assistants to follow complex instructions or handle multilingual tasks?
IBM's Granite-4.0-Micro is here to help enterprises with instruction-following LLMs. This 3B-parameter mini-package brings:
- Accurate summarization & text extraction
- Question-answering & Retrieval-Augmented Generation (RAG)
- Code completions & function-calling tasks
- Multilingual dialog support across 13+ languages

If you’re building AI agents, automating enterprise workflows, or experimenting with advanced LLMs, Granite-4.0-Micro delivers the flexibility and precision you need once you fine-tune or customize it with your own data.
And with NodeShift Cloud, setup, deployment, and scaling are effortless, secure, and GPU-accelerated for enterprises thinking about long term stability.

Here’s a latest demo guide from us for installing & running Granite-4.0-Micro locally:
🔗 Link: https://nodeshift.cloud/blog/get-started-with-ibm-granite-4-0-micro-for-enterprise-rag-summarization-qa-code-tasks?utm_source=telegram&utm_medium=social&utm_campaign=granite_4_micro_launch

What if AI doesn't generate just basic transcriptions, instead understands the audio and describe it with human-level depth?

Meet Qwen3-Omni-30B-A3B-Captioner, a powerful audio captioning model that generates fine-grained, low-hallucination captions across any soundscape.
From multilingual speech and layered emotions to environmental noise, music, and cinematic effects, it delivers detailed, context-aware audio descriptions without requiring extra prompts.

And the best part? With NodeShift Cloud, you can install, run, and start experimenting instantly in a CUDA-ready environment, no complex setup, just speed and scale in minutes.
🔗 Read the full guide here: https://nodeshift.cloud/blog/how-to-install-qwen3-omni-captioner-for-accurate-audio-captioning?utm_source=telegram&utm_medium=social&utm_campaign=qwen3-omni-captioner

Liquid AI just dropped LFM2‑2.6B — a next‑generation hybrid model built for edge AI & on‑device deployment.

With 2.6 B parameters, multiplicative gates + short convolutions, and support for 8 languages, it’s one of the few open models designed to run smoothly on CPU, GPU, and even NPU hardware.

What you can build with it:
✔️ Lightweight tool‑calling agents that work offline or on your laptop
✔️ Data extraction & RAG workflows on private documents
✔️ Conversational assistants with multilingual support
✔️ Creative writing, summarization, etc

What’s inside our new guide
✔️ How to install & run LFM2‑2.6B locally with Transformers
✔️ How to serve it via vLLM for fast, scalable inference
✔️ How to build a minimal agent that calls functions (time, math, RAG) step‑by‑step
✔️ VRAM & GPU tips (BF16 vs. 4‑bit, FlashAttention‑2, sweet spots)

Read the full guide here: https://nodeshift.cloud/blog/pocket-operator-a-local-tool-calling-agent-powered-by-lfm2-2-6b

HUGE RELEASE ALERT!
Qwen team has just dropped a major upgrade of Qwen2.5-VL, the most popular vision model in AI industry, which is used by many big players to fine-tune their domain specific vision models.

The newest version is Qwen3-VL, Alibaba’s new multimodal vision-language model that’s breaking benchmarks and expectations.
We just dropped a full guide on how to install and run Qwen3-VL Locally - step-by-step, clean, and fast.
🧠 Expect next-level multimodal understanding
🎥 Vision + Text synergy
⚡️ Lightning-fast inference with NodeShift

🔗 Read now: https://nodeshift.cloud/blog/how-to-install-run-qwen3-vl-locally-a-step-by-step-guide?utm_source=telegram&utm_medium=social&utm_campaign=qwen3-vl_announcement

IBM launches Granite 4.0-H — a family of long-context, tool-calling LLMs built for real work.

Three sizes, same DNA:
✅ Micro-H (3B, 1M ctx): lightweight & snappy for JSON/IE, routing, short multilingual chat, FIM code.
✅ Tiny-H (7B, 1M ctx): the sweet spot—stronger reasoning, multi-turn assistants, compact RAG, solid tool-calling.
✅ Small-H (32B, 1M ctx): muscle for complex workflows, long-doc comprehension, higher-fidelity coding & analysis.

We just published a hands-on guide to get you productive fast:

What’s inside
✅ Two setup paths: Ollama + Open WebUI (fast chats) & Transformers/vLLM (prod services)
✅ GPU sizing tables for Micro/Tiny/Small + why we standardize on 1×H200
✅ A mini benchmark/prompt pack to compare the three models
✅ Tool-calling scripts (emit/parse <tool_call> and feed <tool_response>)
✅ Minimal Python examples (BF16 & 4-bit) + sanity checks & troubleshooting

Check out the full guide here: https://nodeshift.cloud/blog/how-to-install-run-ibm-granite-4-0-h-tiny-small-and-micro-locally

Big TTS models, running on heavy hardwares, still delivering robotic voices?

Forget them, as NeuTTS Air brings super-realistic, on-device voice AI with instant voice cloning, can run easily on CPUs, no heavy GPUs needed.
- Generate ultra-human voices in real-time
- Clone any speaker in just 3 seconds of audio
- Optimized for laptops, phones & even Raspberry Pis

NeuCodec-powered audio ensures crystal-clear quality with low power consumption.
TL;DR: It’s realistic speech + instant voice cloning + on-device performance, all in one compact model.

In our latest guide, we show you how to install and run NeuTTS Air locally, with NodeShift cloud making setup and GPU-accelerated deployment effortless, get lifelike voice AI running in minutes.
🔗 Read here: https://nodeshift.cloud/blog/how-to-install-and-run-neutts-air-locally-super-realistic-on-device-voice-ai-with-instant-voice-cloning?utm_source=telegram&utm_medium=social&utm_campaign=neutts_air_launch

Meta AI just launched the Code World Model (CWM)!

The Code World Model (CWM) is a 32B parameter dense autoregressive LLM developed by the Meta FAIR CodeGen Team. Unlike traditional code models, CWM was mid-trained on Python execution traces, memory trajectories, and containerized agentic interactions—making it uniquely suited for reasoning about how code affects computational environments.

What’s special about CWM?
✅ Mid-trained on real execution traces & agentic environments
✅ Post-trained with multi-task RL for verifiable coding, math, and multi-turn software engineering
✅ Research-only (non-commercial) release under FAIR license
✅ Strong benchmark performance on Math-500, AIME, and SweBench

We just dropped a full step-by-step guide on:
🔹 Requesting gated access
🔹 Running on a NodeShift GPU VM
🔹 Serving with vLLM
🔹 Streamlit UI for interaction

Read the full guide here: https://nodeshift.cloud/blog/how-to-install-run-facebook-cwm-locally

Still relying on slow, API based TTS that sounds robotic, can't run on small devices and breaks the bank?
Meet KaniTTS, the latest high-speed, high-fidelity voice AI that runs entirely on your device with just some basic GPU acceleration.

What makes it special:
- Powered by a 370M LLM + Neural Audio Codec for ultra-natural, real-time speech
- ~1 sec latency for 15 seconds of audio, perfect for chatbots, assistants & accessibility tools
- Multilingual: English, German, Chinese, Korean, Arabic & Spanish
- Runs locally with just 2-4GB GPU memory, no APIs, no data leaks, no lag

With NodeShift cloud, setting it up is effortless, GPU-optimized, ready-to-run, and privacy-first.
Get studio-quality speech generation right on your own hardware in minutes.
🔗 Read the full guide: https://nodeshift.cloud/blog/how-to-install-and-run-kanitts-locally-real-time-on-device-voice-generation?utm_source=telegram&utm_medium=social&utm_campaign=kanitts_launch

The ModernVBERT team has just unleashed a compact 250M-parameter vision-language model, which is matching the performance of models up to 10x larger, performing way above its weight!

With state-of-the-art multimodal reasoning, advanced document retrieval capabilities, and seamless image + text understanding, ModernVBERT is your go-to model for next-level AI & RAG workflows.

We’ve published a step-by-step guide to install and run ModernVBERT locally - fast, clean, and ready for experimentation.
- Unlock multimodal intelligence
- Advanced visual document comprehension
- Optimized for lightning-fast local inference with NodeShift Cloud

🔗 Read the full guide here: https://nodeshift.cloud/blog/how-to-install-modernvbert-compact-vlm-for-document-retrieval-in-rag-applications?utm_source=telegram&utm_medium=social&utm_campaign=modernvbert_launch

ServiceNow just released: Apriel-1.5-15B-Thinker

Apriel-1.5-15B-Thinker is an open-weights, multimodal reasoning model (image-text-to-text) focused on strong mid-training/continual pre-training plus high-quality text SFT—no RL required. It’s compact (15B) yet competitive with much larger models and designed to run on a single GPU.

We just published a step by step guide to install and run Apriel locally—plus a simple Streamlit UI so you can chat with the model and ask questions about images.

What the guide covers:
✅ Picking a GPU + VRAM sizing tips
✅ CUDA/PyTorch install (cu121) & env setup (Py 3.11)
✅ One-file for text + vision with the correct dtype cast (BF16/FP16)
✅ Optional Streamlit app (text & image tabs, sliders for temp/tokens)
✅ Tuning for speed/VRAM (token limits, fp16, 8-bit options)

Check out the full guide here: https://nodeshift.cloud/blog/how-to-install-run-servicenow-apriel-1-5-15b-thinker-locally

Microsoft just released UserLM-8B — a unique open-weights model that flips the script: instead of acting like an assistant, it simulates the user in a conversation.

What is UserLM-8B?
A fine-tuned Llama-3.1-8B model trained on WildChat-1M to generate realistic user turns—from the first query to multi-turn follow-ups—and it can even gracefully wrap up with a special <|endconversation|> token.

Why it’s special
✅ Purpose-built for assistant evaluation & robustness testing
✅ Great for synthetic dialogue data generation
✅ More natural, diverse “user” behavior vs. prompting an assistant model to pretend

We just published a new guide “How to Install & Run Microsoft UserLM-8B Locally”

What’s inside:
✅ GPU sizing + a practical VRAM table
✅ Full setup on a GPU VM (NodeShift example)
✅ Ready-to-run scripts
✅ Guardrails for realistic simulations (stop tokens, end-of-conversation handling)
✅ Tips to plug UserLM into your own assistant for end-to-end testing

Check out the full guide here: https://nodeshift.cloud/blog/how-to-install-run-microsoft-userlm-8b-locally

Wondering how can you test if your AI model is safe or behaves unpredictably under pressure?
Anthropic has just released a non-negotiable tool for AI safety - Petri (Parallel Exploration Tool for Risky Interactions). Petri is an open-source tool that automates AI behavior testing through multi-turn conversations, simulated environments, and detailed scoring across safety dimensions.

With Petri, you can:
- Run alignment tests on any model - deception, reward hacking, & more
- Automate hundreds of behavioral evaluations in minutes
- Get structured insights and transcripts for deeper analysis & benchmarking

And with NodeShift Cloud, you can install and run Petri locally, easily, securely, and with zero setup friction.
In our latest guide, we’ll cover:
🔹 How to install and set up Petri locally
🔹 How to setup local model for auditing with Ollama as the API
🔹 How to run your first automated safety audit
🔹 How to provide seed instructions and interpret transcripts

🔗 Read full guide here: https://nodeshift.cloud/blog/how-to-install-run-anthropics-petri-locally-the-easiest-way-to-audit-ai-models-for-safety?utm_source=telegram&utm_medium=social&utm_campaign=petri_ai_audit

Qwen just dropped: Qwen3-VL-30B-A3B-Thinking

A powerhouse multimodal model built on a Mixture-of-Experts stack—designed for deep text + vision + video reasoning, long-context understanding (256K→1M), robust OCR (32 languages), GUI/tool use, and even converting diagrams/screens into working code.

We’ve published a fresh, hands-on guide:
“How to Install & Run Qwen3-VL-30B-A3B-Thinking Locally” — tuned for a GPU VM workflow (We used NodeShift, but it works anywhere).

What’s inside the guide
✅ Clean environment setup (CUDA-aligned PyTorch, optional FlashAttention-2)
✅ Image & video inference
✅ “Thinking” variant notes + practical VRAM plans (single-/multi-GPU)
✅ Troubleshooting (FA2 mismatches, SDPA fallback)
✅ Ready-to-copy commands & code blocks for Jupyter/terminal

Check out the full guide here: https://nodeshift.cloud/blog/how-to-install-run-qwen3-vl-30b-a3b-thinking-locally

AI21 Labs just launched Jamba Reasoning 3B — a compact, hybrid Transformer–Mamba model built for serious reasoning on modest hardware.

Why it’s special
✅ ~3B params (26 Mamba + 2 Attention) → fast, memory-light, edge-friendly
✅ 256K context without the usual KV-cache blow-up
✅ Strong benchmarks: IFBench 52.0, Humanity’s Last Exam 6.0, MMLU-Pro 61.0
✅ On-device speed that holds up as context grows (≈43–44 tok/s at 16–32K)

We just published a new step-by-step guide:
“How to Install & Run AI21-Jamba-Reasoning-3B Locally (GPU VM)”

What’s inside
✅ Pick the right GPU & VRAM (rule-of-thumb table)
✅ Clean setup on a CUDA 12.1.1 image (Python 3.11, Torch cu121)
✅ vLLM serving (OpenAI-compatible) with the right flags for Mamba SSM
✅ Transformers alternative path + FlashAttention 2 tips
✅ A one-file Streamlit UI to chat with the model on your own server

Check out the full guide here: https://nodeshift.cloud/blog/how-to-install-run-ai21-jamba-reasoning-3b-locally

VNGRS Releases Kumru-2B — A Turkish-Native Lightweight Language Model

VNGRS has officially released Kumru-2B, a compact yet powerful Turkish-native LLM built entirely from scratch. Trained on ~500 GB of curated text (≈300B tokens) and fine-tuned on over 1M supervised examples, Kumru-2B is designed specifically for the Turkish language — featuring a modern 50K-token Turkish-optimized tokenizer, 8K context window, and native support for math and code.

Why Kumru-2B is Special
✅ Built from scratch for Turkish — not a multilingual adaptation.
✅ Efficient tokenizer: uses ~40% fewer tokens than multilingual models like GPT-4o or Gemma.
✅ Punches above its weight — outperforms much larger models like Llama-3.3-70B and Qwen2-72B on Turkish-centric tasks.
✅ Runs smoothly on local or cloud GPUs, making it ideal for research, startups, and developers.

In our latest blog, we walk you through everything you need to:
✅ Deploy a GPU-powered VM on NodeShift Cloud
✅ Install Python 3.11 + CUDA 12.1.1 environment
✅ Run the model with a simple Python script
✅ Launch an interactive Streamlit WebUI to chat with Kumru-2B directly in your browser

Whether you’re building NLP tools, studying Turkish linguistics, or experimenting with LLMs, this guide helps you get started in minutes.

Read the full tutorial here: https://nodeshift.cloud/blog/how-to-install-run-vngrs-ai-kumru-2b-locally