Open-source AI software and open-weight models you can run yourself.
Recursive's Flint is a 40B MoE coding model with 2.5B active parameters, released under Apache 2.0. Here is what it scores in Recursive's own tests, how fast it runs on a Mac or RTX Spark, the hardware you need, and what the Eureka Machine built.
Bria's Fibo 2.0 is an 8B model that edits and generates images from structured JSON prompts, trained on licensed data. Here is how its five modes work, Dev vs Edit Turbo, what Bria claims on benchmarks, and why the weights are non-commercial while hosted use costs $0.04 per image.
DeepSeek open-sourced DeepInfer, the Apache 2.0 inference engine behind DeepSeek-V4.1-Flash, for NVIDIA SM100 and Huawei Ascend 950. What it includes, what it leaves out, and what DeepSeek's claim of 155 trillion tokens a day and $38 per GPU-hour does and does not show.
Asymptote Labs raised a $5.5M seed for Beacon, an MIT-licensed tool that records sessions from Claude Code, Cursor, Codex and 20+ other agents and turns reviewed lessons into reusable skills. How it works, what Pro costs, and what to check first.
Nvidia's Nemotron 3.5 Bolt 74B-A9B packs a model compressed from Super VL into 24GB with 2-bit experts for RTX Spark and DGX Spark. Here is what it needs to run, how its speed compares in Nvidia's own tests, and when Bolt E4B arrives.
The best open-weight LLMs as of October 2026, sorted by hardware: GLM-5.3, Kimi K3, DeepSeek V4.1-Flash, Muse Spark Open and Beam for servers; Nemotron 3.5 and GLM-5.3-Flash for one node; Qwen3.8-27B, Gemma 4, LFM2.5 and MiniCPM-V 4.7 for your own computer.
Reflection's 501B open-weight Beam scores 30 on Artificial Analysis's index and 23.5% on ARC-AGI-2 at $0.61 a task. X users praise the Apache 2.0 license and token efficiency, question a 28 coding-agent score and Qwen 3.8 27B comparisons, and note it needs about 280 GB even at 4-bit.
Alibaba's Qwen3.8 covers a 2.4T-parameter API flagship, its first Max-class open weights, a 27B Apache 2.0 model, Flash at $0.15 per million input tokens, omni-modal models, and Flash-Next, a preview of the Qwen 4 architecture. Which model is which, what each costs, and what is next.
A beginner's guide to running an AI model on your own computer with Ollama, LM Studio or llama.cpp: hardware needs, install steps, small models to start with, and our real test results on a CPU-only Linux machine, as of October 2026.
inclusionAI, Ant Group's AGI lab, released Ming-3.0-audio-flash (126B, 6.5B active) and Ming-3.0-audio-tiny (9.3B) under MIT. They transcribe, reason over audio and talk back. Vendor charts show leading speech recognition but gaps on meetings and long audio.
MOSS-Image is a new Apache 2.0 text-to-image model from OpenMOSS built for dense text and layouts. Here is what its benchmarks show, how its JSON layout prompts work, what it cannot do yet, and how to run it in Diffusers or ComfyUI.
Mistral's Timestral forecasting models run on a plain CPU, and one tester reproduced 14 fev-bench scores. Hacker News and X users praise the small 28M model, question a thin leaderboard lead, and push back on a weights license that bars companies above $20M in monthly revenue.
OpenBMB's MiniCPM-V 4.7 is an Apache 2.0 vision model in 35B-A3B and 1B sizes. A new expert-scheduling method streams MoE weights from flash, and OpenBMB reports up to 20.51 tokens per second on an iPhone. Here is how it works, how it scores, and how to run it.
Santander's AI Lab says smaller, specialized models cut serving costs from around €5 to about €0.03 per interaction in selected experiments. Here is what it built with NVIDIA, what the three-year deal covers, and what the numbers do not show.
EdenLive lets a game engine or coding agent run the world while a 5B or 14B open model repaints its rough proxy video in real time. VAST reports 29.8 FPS at 1280 x 704 on one H100 and about 24.7 FPS on an RTX 5090. Here is how it works, what it needs, and what the license allows.
Ant Group's inclusionAI has published Ling-3.1-flash, a 560B MoE with 25B active parameters, under an MIT tag. Its own charts put it behind Claude Opus 5 on SWE-bench Pro but ahead on HealthBench Professional. Self-hosting needs about 574 GB of FP8 weights and eight H200s.
Black Forest Labs' FLUX 3 Image lets you place elements with bounding boxes, edit one part of an image, use up to 10 references and render in 4K. Klein, the 5.6B open-weights version, starts at $0.008 per image, and its INT8 build runs on RTX GPUs with 8 GB of VRAM.
Free datasets for machine learning, grouped by data type: tabular starters like Iris and Adult, images from MNIST to Open Images, text, speech, LLM-scale web data, and public data. Sizes from each dataset's own page, plus the license to check before you ship.
Google DeepMind released EmbeddingGemma 2 on October 6, 2026: a 740M-parameter Apache 2.0 multimodal embedding model for text, code, images, video, and audio. Specs, benchmarks, MRL truncation, and how to run it.
Mistral Large 4 is a 1.05T-parameter multimodal MoE with 49B active parameters. It is in API preview at $1.36/$4.18 per million tokens, scores 38 on the Artificial Analysis index, and its open weights are due by October 31. The license is still unpublished.
