Videos
Find videos about LLM models, hardware, runtimes, benchmarks and practical AI workflows.
Demonstrates an open-source Mission Control dashboard for managing an AI software factory, showing version comparisons, decision routing, and integration with Archon and Neon Postgres.
Benchmarks Qwen3.8 27B and Flash Next on one/two Intel Arc Pro B70 GPUs via llama.cpp, vLLM, Strata, comparing prompt processing and generation against Radeon AI PRO R9700 and Strix Halo.
Evaluates running mixture-of-experts models from 4 GB to 16 GB VRAM using llama.cpp, culminating in DeepSeek V4.1 Flash on one card, with benchmarks via Artificial Analysis Intelligence Index v4.3.2.
Evaluates whether a locally-hosted model can compete with Kimi K3.1 Pro, described as the largest open-weights model at 2.8T parameters, through challenges and code reviews.
Raon 1 Mini is evaluated via Performance, Memory, Reasoning, HumanEval Remix, Kanban, Sand Physics, Dungeon Crawler, Blender, and Godot tests using RTX 2000 and RTX 5090 hardware.
Demonstrates Colibri v2.1 installation, system configuration, concept diagrams, chat UI, monitoring, profiling, running Flint GGUF, and testing Qwen3.6-35B-A3B and DeepSeek V4 Flash on one GPU.
Locally installed SenseNova-Loop-9B-Preview is tested on handwritten equations, beam diagrams, invoice to JSON conversion, text to SQL generation, and workflow diagram creation.
Demonstrates running Liquid AI's LFM3 Text and LFM3 Omni models locally on Apple Silicon via oMLX Ask AI for real-time voice conversations and text chat.
Local installation and testing of MOSS-Image-SFT using resources from OpenMOSS-Team/MOSS-Image-SFT on Hugging Face.
Tests quantized Mistral Large using Inferencer v2.5.1 on M3 Ultra 512GB across benchmarks, coding, logic, creative tasks, and compares Local vs Cloud performance.
Analyzes Anthropic's study showing analysts followed agent recommendations 86% of the time using their wording, demonstrating influence on expert judgment via Claude Code examples.
Hands-on local tutorial demonstrating how MeshLLM pools GPUs and memory across multiple machines into one unified system using the mesh-llm GitHub repository.
Eight AI video models—Kling 4.0, Seedance 2.5, Gemini Omni Flash 1.1, FLUX.3 Video, Grok Imagine 1.5, Wan 3.0 Prime, MiniMax H3 and Seedance 2.0—are compared using identical prompts on Higgsfield.
Locally installed Surya 2.5, a 689M parameter OCR model, is tested on handwritten French letters, old Polish passports, invoices, tables, an 1883 newspaper masthead, LaTeX pages, Hindi, Arabic, Urdu,
Evaluates Kling 4.0 on Higgsfield using pre-written unit tests across References, Frames, and Edit video modes, reporting six out of eight tests passed for 30-second clips.
Locally installs and tests Phonon-2, described as the most accurate open speech recognition model for English under 900 MB.
Tests InternScience/Sage-A1.5-Q4_K_M-GGUF on RTX 2000 and RTX 5090 using Performance, Memory, Reasoning, HumanEval Remix, Seat Planner, Sand Physics, Dungeon Crawler, Blender, and Godot benchmarks.
Locally installed Sage-A1.5 undergoes testing via Jupiter gravity assist simulation, parallel subagent orchestration, and vision-based forensic physics analysis on photos.
Demonstrates Qwen3.8-Flash-Next controlling a Seeed Studio reBot Arm B601 via plain-English instructions for throwing, shooting, writing, drawing and pouring, using Meta SAM 3.1 for vision.
Locally installs and tests NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, an open-weight, offline speech and audio model.
Compares updated Strata Coder and Turbo Brilliance via 9 tests: speed, strategy game, tool calling, CAPTCHA, 262K context, SVG, voxels, motion design and physics puzzles.
Demonstrates local execution of Liquid AI LFM3-1.7B on an Apple Mac mini M6 with 32GB unified memory via OpenClaw, achieving approximately 70 tokens per second even at high context utilization.
Evaluates Gemini 4 Carbon against Claude Opus 5.5 and GPT-6 Astra via benchmarks, reviews Fable 5.5 leaks, Obsidian 3, Claude Ultrathink, and covers updates from OpenAI, NVIDIA, Google, and Apple.
Locally installs and tests Clover-35B-A3B-Preview, evaluating performance through Wildfare bug fixes, animations, multilingual tasks, benchmarks and zebra logic puzzles.
Demonstrates building an AI second brain using Context, Connections, Capabilities and Cadence via tools and markdown files, referencing Nate Herk's AIS-OS framework.
Runs Inferencer v2.5.0 with DeepSeek V4.1 SE and Qwen 3.8 Flash SE on M1 Max, M5 Max and Mac Studio M3 Ultra hardware.
Genie 4 by Google unlocks unprecedented realism and improves real-time interactive simulation capabilities within its world model framework.
Introduces Gemini 4 Carbon, Google's strongest model featuring state-of-the-art cybersecurity and long horizon capabilities.
Evaluates Pocket TTS running entirely on CPU with no VRAM and Ming 3 TTS featuring high-quality voice cloning with paralinguistic tags through demonstrations within license limits.
Tests LiquidAI LFM2.5 24B A2B against Qwen 3.6 35B A3B using Performance, Memory, Reasoning, HumanEval Remix, Kanban, Sand Physics, Dungeon Crawler, Blender and Godot benchmarks.
Tests Spark-X2.6-4B performance in coding, reasoning, tool calling, and agentic AI workflows on Apple Mac mini M6 with 32GB unified memory via LM Studio and Pi Coding agent.
Tests Liquid AI LFM3 1.7B with Hermes agent workflows, plus vision, voice, and omni capabilities, evaluating performance on edge devices including phones.
LFM3-1.7B is run locally via llama.cpp and tested on instruction following, structured output, multilingual tasks, reasoning, hallucinations, grounded answers, and tool calling.
Locally installed Unlimited OCR 2 parses handwritten physics equations, mixed-language handwriting, invoices, tables, printed math solutions, and an 1883 newspaper ad using SR-SWA.
Tests building a functional game asset editor using Ternary Bonsai 2 27B within 12GB VRAM constraints, comparing its capability against Qwen 3.8 27B through iterative development stages.
Demonstrates integrating SmartBear BearQ APIs with GitHub Actions using QA Lead, Explorer, and Tester Agents to autonomously execute regression tests upon code pushes or pull requests.
MiniMax M3.1 Flash and Claude Opus 5.5 were given three tasks from a 20,000-line Go project, passing every hidden check with costs of approximately $0.65 and $7 respectively.
Tests GPT-6.1 Terra on browser workflows, C++ game dev, robot arm control, FPS gen, Blender, Godot, creative writing, and computer use via Mensa and Blender tasks.
Locally runs DeepSeek-V4.1-Flash-1017 GSQ-RCO GGUF using Strata on 12 GB VRAM, explaining non-uniform quantization methods and comparing full versus quantized weights.
Benchmarking Kimi K3.1 Pro Standard, High, and MAX modes against DeepSeek and local Kimi K3 using agent workflows, coding challenges, logic riddles, and simulations.
Installs and tests Nanbeige4.3-33B using code debugging and papel picado canvas tasks, alongside benchmarks evaluating its performance as a lightweight agent.
Tests running DeepSeek v4.1 Flash using Strata to achieve operation in 12GB VRAM, comparing performance against Qwen through coding tasks.
Tests Flint by RecursiveSI using Performance, Memory, Reasoning, HumanEval Remix, Kanban, Sand Physics, Dungeon Crawler, Blender and Godot on RTX 2000 and RTX 5090.
Tests DecisionNode via script, CLI, point-and-box tasks, prompt injection resistance, custom rules, and Python SDK batch processing using text, JSON, or image inputs.
FLUX 3 Image Klein is installed with ComfyUI and tested using Nemotron 3.5 Super VL for tasks including multilingual translation, circuit diagrams, and HTML generation.
Tests Strata Coder IQ1_M against Qwen 3.8 Flash Next NVFP4 using 9 benchmarks: speed, battle strategy, tool use, visual CAPTCHAs, long-context retrieval, SVG, voxels, motion and physics.
Reviews Dessa, an AI developed by Block that learned to talk by listening.
Tests NVIDIA-Nemotron-3.5-Bolt-E4B-4GB using the Raramuri Runners Test, Yakisoba Teppan HTML Simulation, logical puzzles, science reasoning, scheduling tasks, and baby names in 82 languages.
Kimi K3.1 Pro, Claude Opus 5.5 and GPT-6 Sol each have one hour to independently write code, run, debug and deliver a playable 3D racing game.
Fine-tunes Falcon-H1-Tiny 90M via MLX LM on MacBook Pro M5 Max to create a prompt injection firewall, achieving 96.9% attack detection and 7.4% false positive rate, then deploys in Ollama.