Home

Open Source

Curated open-source AI projects with verified GitHub data and hands-on breakdowns.

5.6B Beats 9B: FLUX 3 Klein Opens Its Weights

Black Forest Labs put FLUX 3 Image Klein's weights on Hugging Face on 2026-10-15, at huggingface.co/black-forest-labs/flux-3-image-klein, alongside a fine-tunable base repository and a gated ComfyUI repository. The image generator is a 5.6B flow-matching transformer built, per the official description, from per-mode early blocks plus 28 joint blocks, with a Qwen3-VL-4B text encoder and the existing FLUX.2 VAE; parameter counts are each publisher's own and exclude text encoders. The headline is that the smaller generation wins: against its own predecessor FLUX.2 [klein] 9B it gains 91 Elo in text-to-image and 70 in image editing with 38 percent fewer parameters, and in BFL's internal evaluations it ranks first in text-to-image, multi-reference generation and image editing, ahead of the closed Ideogram V4.5 and its own FLUX.2 [pro], with Qwen-Image 2.1 at 7B and Krea 2 at 12.9B as the open comparisons. Krea 2 does not support image editing and so is absent from that comparison - a detail worth keeping, or three firsts get misread as sweeping every rival in all three. Capabilities: native 4K with dimensions in multiples of 16, pixel-perfect editing where untouched regions are preserved exactly rather than approximately repainted, bounding boxes that place elements or confine edits to a region, up to 10 reference images for character consistency and style transfer without retraining, and an optional Qwen3.5-9B prompt upsampler. Memory comes in three tiers: 11.2GB in BF16, 5.6GB in FP8, and an INT8 build that runs on NVIDIA GeForce RTX cards with as little as 8GB of VRAM through ComfyUI or Diffusers. Vendor-reported speed puts a 1MP image at about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 with the distilled INT8 build, under 0.6 seconds in NVFP4 on an NVIDIA GB300, with a served LoRA adding roughly 5 percent. The licence is the FLUX Kommunity License: free commercial use for individuals and companies under 5 million dollars in annualized revenue, free research and non-commercial use for everyone, and self-hosted licensing or the API beyond that threshold, with API pricing from 0.009 dollars per 1K image, the LICENSE text being the final authority. Three caveats belong side by side: every quality figure is BFL's internal evaluation with no independent reproduction yet; the API latency comparison is not like-for-like, since BFL timed its own servers while rivals were timed from the client including polling and network; and the 8GB claim is not tied to a specific variant, output size or generation time, with the published speed charts not testing an 8GB configuration or 4K output at all. Day-one support covers ComfyUI via PR 17064 plus four official templates, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni, with LoRA training through Ostris AI Toolkit and NVIDIA NeMo AutoModel.

A 60 FPS World Model on One RTX 5090. Its Own Site Shows 0

SkyworkAI, a unit under Kunlun Wanwei, pushed Matrix-Game 4.0 to GitHub on 2026-10-15 (the series repo github.com/SkyworkAI/Matrix-Game was created 2025-05-06, with 4.0 living in the Matrix-Game-4/ subdirectory and a homepage at matrix-game-v4-0.github.io). A GitHub API snapshot on 2026-10-16 shows 2,488 stars and 272 forks - the aggregate for the series repo, folding in versions 1.0 through 3.0, so it must not be read as 4.0's own growth. The core idea: the game engine first renders a low-poly proxy frame, then a causal DiT video model performs streaming neural rendering into the final image. Three layers: a game engine holding state, physics and input; a coding agent that turns natural-language instructions into new rules that rewrite the engine while retrieving engine state to build a dynamic prompt; and a causal DiT distilled to one or two denoising steps where each latent yields four frames. HUD and sound are frame-aligned presentation layers outside the model. Three scenes - Medieval Castle, Ancient Egypt and City Racing - with weights and assets on Hugging Face as Skywork/Matrix-Game-4.0. The license must be read in three layers, LICENSE originals as authority: the root LICENSE is MIT (Copyright 2025 SkyworkAI and contributors, covering older series code); the v4 directory's Matrix-Game-4/LICENSE is the full Apache-2.0 text (11,357 characters); and inside v4's bundled licenses/ directory, MG4CF keeps Apache-2.0 while TAEHV keeps MIT (Copyright (c) 2025 Ollin Boer Bohan). The README states it directly: project code is Apache-2.0, vendored components retain their original licenses, and model and asset distributions carry their own statements. The Hugging Face model card's license field could not be retrieved for this piece because the lookup was blocked, so the article writes only that the model card is the final authority and does not guess a license name. Three sets of performance figures must be written side by side: (1) official launch figures of a single RTX 5090, 480p at 60 FPS, end-to-end latency under 150 ms, an 83 FPS steady single-step denoising rate and 18.5 GiB peak VRAM, not independently reproduced; (2) the project page performance panel is still a placeholder at 0 FPS, 0 ms, 0p and 0 GPU, with a footer saying content will be updated at release, visited on 2026-10-16; (3) the repository README's own measurements, the hardest of the three - one A100-SXM4-80GB at 864 x 480 with one denoising step plus TAEHV, SDPA, ring and fused DiT and CUDA graphs, giving about 38-39 FPS for the two shooter scenes with fixed prompts (model processing within an online session), about 38.5 for racing fixed and about 28.9 for racing dynamic (server-side online pipeline throughput), with the README stating that --fps 40 is a target capture cadence rather than a throughput or latency guarantee, that these are neither browser presentation FPS nor action-to-photon latency, and that the racing rows come from an earlier checkpoint. Version 4.0 also ships RTX 5090 presets (--preset rtx5090 on PyTorch SDPA, rtx5090-sage on SageAttention) that convert DiT block linear weights to FP8 at load time and quantize activations at inference while keeping embeddings and the output head in BF16, plus fused kernels and CUDA graphs; a version for the RTX 5080 and 5070 is officially slated for next week. Boundaries worth knowing: one server process supports only one active session and one scene model; networking binds to loopback by default under a security statement calling this a trusted single-user demo rather than an authenticated public cloud-gaming service; and the legacy Windows entry is retained but not validated for feature parity. Repository address: github.com/SkyworkAI/Matrix-Game.

On-Device AI: 3GB RAM Runs a 35B MoE Model

ModelBest and the Tsinghua THUNLP-backed OpenBMB team released MiniCPM-V 4.7 (official repo changelog October 12, 2026; press coverage October 13), billed officially as the first open MoE multimodal model to run losslessly and smoothly on-device, in 35B-A3B (35B total / 3B active per token) and 1B variants. Speed per the official documentation only: 13+ tokens/s at about 3 GB of memory on an iPhone 17 Pro Max (third-party 15+/17+ figures conflict with the docs and are not adopted); the 3 GB is the model's runtime footprint, not the device floor - expert weights stay on flash and stream in on demand (MoE-SSD), and the official app hardware table recommends 12 GB+ RAM for 35B-A3B and 4 GB+ for 1B. The core trio: LES (Lookahead Expert Scheduling - cross-layer routing predictions affect prefetching only, never the compute routing, so misses cost nothing in correctness; officially a 5x real-device speedup, claimed as the industry's first open-source implementation), fine-grained layered quantization, and DSpark speculative sampling; vision keeps hybrid 4x/16x token compression and adds intra-ViT early compression based on LLaVA-UHD v4. The ecosystem: FP8/int4/AWQ/GPTQ/GGUF weights on Hugging Face and ModelScope; inference via SGLang, vLLM, llama.cpp and Ollama; fine-tuning via SWIFT and LLaMA-Factory plus the co-open-sourced SRFT self-replay method; the official inference path runs on OpenBMB/llama.cpp-omni (MIT, a fresh llama.cpp-based repo created October 13); official apps for iOS, Android and HarmonyOS live in the sibling repo OpenBMB/MiniCPM-V-Apps; the Day-0 chip list per the docs is Intel, AMD, Arm, Rockchip and Houmo (FlagOS adds Ascend, Moore Threads, MetaX - claimed). License: weights and code are unified under Apache-2.0 (docs/project_en.md verified: "weights and code are open-sourced under the Apache-2.0 license"); the registration questionnaire is optional and not a commercial precondition. GitHub API snapshot October 15, 2026: 26,753 stars / 2,096 forks, pushed October 14. Repo: github.com/OpenBMB/MiniCPM-V.

Agent Skills' npm Moment: 34,494 Stars, One Prompt CLI

vercel-labs/skills - "The open agent skills tool - npx skills" - turns reusable instruction bundles kept in Git repositories (skills) into installable, updatable, distributable dependencies; media call it npm for agent skills. Verified via GitHub API on October 14, 2026: 34,494 stars / 2,952 forks, MIT (LICENSE file confirmed: Copyright 2026 Vercel, Inc.), TypeScript, created 2026-01-14, pushed October 9, on the daily trending board. It covers 79 coding agents (Claude Code, Codex, Cursor, OpenCode and more); the command set is add/use/list/find/remove/update/check/init; installs default to project-scoped symlinks (--copy fallback, -g global, --all -y unattended CI); the use command tries before installing (temp directory, prompt printed to stdout); sources include GitHub shorthand, GitLab, Azure DevOps, arbitrary git URLs, local paths and private-repo credentials; limits are 10MB download / 25MB extraction / 1000 files (env-overridable); the skills.sh directory is populated by anonymous install telemetry (disable with DISABLE_TELEMETRY=1 or DO_NOT_TRACK=1). An ecosystem route debate: Cloudflare filed a competing .well-known/skills proposal (RFC 8615). The official collection vercel-labs/agent-skills (32,367 stars; no license marked via the GitHub API and no LICENSE file found at the repo root - this article makes no license claim for it) includes react-best-practices (40+ rules in 8 tiers), web-design-guidelines (100+ rules) and more. Real-world adoption shows up in suggestion data (install commands for wecomteam/okx/tencent and others). Honest limits: a young project created in January 2026; skill quality depends on author and attached model; symlinks break in some environments. Repos: github.com/vercel-labs/skills and github.com/vercel-labs/agent-skills.

God's Eye View: A Real-Data Spy Satellite Simulator

God's Eye View (bilawalsidhu/gods-eye-view) bills itself as a spy-satellite simulator in your browser where the data is real: a CesiumJS WebGL photorealistic 3D globe overlaid with live aircraft (ADS-B), global vessels (AIS), 838 tracked orbital objects (CelesTrak TLE via SGP4), earthquakes, roughly 3,600 public cameras, radio stations, wildfires, launches and weather layers (layer counts mostly per project docs and third-party reports). Verified via GitHub API on October 13, 2026: 53,844 stars, 10,627 forks, JavaScript, created 2026-06-22, pushed October 12; No. 2 on the daily trending board that day (+1,895 stars), after topping daily and weekly boards in August (README self-reported; Trendshift independently logs August 27 as No. 1). Author Bilawal Sidhu is a former Google product manager. License verified from the LICENSE file: MIT (Copyright 2026 Bilawal Sidhu; the GitHub API license field says NOASSERTION, which is wrong here). Practical notes: 13 of 15 layers need no API key (third-party figure); the base layers work keyless; photorealistic 3D needs a Cesium ion key (personal non-commercial terms) or Google Maps; the voice AI agent needs an OpenAI key; install via git clone + npm ci + npm run doctor + npm run dev (Node 18+) or Pinokio. Known traps: old versions hit the public OSM Overpass endpoint and get rejected (update before expecting traffic layers), and dense Cesium entity streams plus 3D Tiles eat VRAM. Honest limits: it is a geospatial data-fusion client, not live satellite imagery; traffic is simulated along real roads; CCTV poses and rocket trajectories are coarse estimates. Sits alongside our September teardown (32,399 stars): this is the October hands-on view. Repo: github.com/bilawalsidhu/gods-eye-view.

4GB VRAM Holds 2.8 Trillion Parameters

AirLLM (lyogavin/airllm) takes the opposite approach: instead of quantizing, distilling or pruning weights, it streams them layer by layer into VRAM, trading speed for capacity. Verified via GitHub API on October 12, 2026: 36,747 stars, 3,829 forks, Apache-2.0, created 2023-06-12, pushed 2026-10-12, 160 open issues, primary language Jupyter Notebook - which makes it more of a research tool than a production component. Repo: github.com/lyogavin/airllm. README measurements (vendor self-reported, not reproduced here): Qwen3.8-27B dense VL in 3.33GB (RTX 3090 end-to-end, needs transformers 5.8+); Qwen3.8-Flash-Next 125B in 5.95GB (RTX 4090, n-gram table file-mapped to host, ~64GB-class RAM, ~360GB checkpoint); Kimi K3 2.8T in 3.72GB (RTX 6000 Ada, per-expert streaming by token routing, requires compressed-tensors + flash-attn + a CUDA 12 torch build + transformers 4.56.x); DeepSeek-V3 671B around 12GB; training support since September 2026 runs 125B under 6GB (RTX 3060 Ti). Honest limits: the headline 70B-in-4GB claim carries no hardware; the README publishes no tokens-per-second figure, so none is invented here; layer-wise streaming means interactive token-by-token use stays awkward; disk and host RAM are the real gates; and Kimi K3 weights use a modified MIT license, not the standard one - read the LICENSE before commercial use. Positioning: our existing tool comparison covers deployment tools, this piece covers the memory-strategy dimension.

AI Agents Can Now Reverse Engineer Apps: 84k Stars in a Week

morluto/rea (Reverse Engineer Anything) topped GitHub Trending: verified via GitHub API on October 11, 2026 - 84,088 stars, MIT, TypeScript; the star-growth chain cites three sources separately (~10,400 via trending data, 47,884 via a third party quoting the GitHub API, 84,088 verified by us). Repo: github.com/morluto/rea. It is an MCP server plus CLI (npx rea-agents setup) that wires your coding agent (Claude Code/Codex/Cursor/Gemini CLI and more) into a reverse-engineering toolbox: native binaries through Hopper/Ghidra/IDA, static JS/Electron with no engine needed, .NET assemblies, APKs (JADX), firmware (Binwalk/Unblob), EVM bytecode, HAR/mitmproxy captures and websites. v6.0 (release notes, October 8, 2026) added pwntools offline ELF layout, crash inspection, Windows PE x86 and LLDB observation, with v6.1-v6.3 following within 24 hours. Honest limits: it does not recover original source, does not clone apps wholesale, and every conclusion ships with evidence and stated caveats; analysis runs locally so the target binary never leaves your machine. README-cited showcases: DX-Ball function rebuilt in C passing 3,205 original x86 cases with all 63 bytes reproduced; the Notion clipboard bridge traced end to end. created_at 2026-04-14 (API) vs a reported October 7 launch - both stated, no exact date asserted.

From 2% to 47%: JetBrains' RL Bet on Mellum2.1

JetBrains shipped Mellum2.1 on October 8, 2026 (HF verified via hf-mirror: ungated, Apache 2.0, early-stage 198 downloads/77 likes). Repo: huggingface.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking. Architecture is identical to June's Mellum2 (12B MoE / 2.5B active / 131K context); what shipped is a training recipe - RL moved from a finishing stage to the main pipeline across millions of sandboxed runs in thousands of environments, with new math/competition/science/tool-use/software-engineering tasks and filtered sources. JetBrains self-reported benchmarks (Pi harness v0.73.1, thinking mode, no third-party rerun): SWE-bench Verified jumped 2.0 to 47.0, but Qwen3.5-9B sits at 50.0 - Qwen wins real-repo long tasks while Mellum wins LiveCodeBench v6 (82.0) and BFCL v4 (62.3); cite each vendor's own table, most figures exist only as charts. Speed: "almost twice Qwen3.5-9B throughput under heavy load on H200" is an official claim; the 1.6x single-request gain rides on MTP speculative decoding and the MTP head was not released with the weights (coming soon) - you cannot buy that speedup today; GGUF (llama.cpp/Ollama/LM Studio) is also coming soon. Deployment: vLLM supported today (reasoning-parser qwen3), BF16 weights around the 24GB class (estimate), no hosted API. Positioning: agent density on a fixed GPU budget with code that never leaves your servers - not a blanket Qwen beater.

Open Video Model LTX-2.5: Diffusion Decoder, 12GB VRAM Entry

Lightricks shipped LTX-2.5 on August 11, 2026, replacing LTX-2.3. Repos (dual form): huggingface.co/Lightricks/LTX-2.5 (HF API verified October 9, 2026: gated=auto, meaning log in and accept terms for automatic access; license tag "other" = the custom LTX-2.x Community License; 1.689M downloads; the Diffusers variant Lightricks/LTX-2.5-Diffusers is also gated=auto), code at github.com/Lightricks/LTX-2. Architecture: a 22B audiovisual backbone plus a custom Gemma 4 12B text encoder and a diffusion video decoder - the VAE decoder is itself a diffusion model, so aggressive latent animation compression and high fidelity no longer trade off; that is this generation's headline architectural change, plus DFR adaptive compute allocation. Capabilities: 4K (3840x2176) HDR, up to 50fps, 24kHz stereo generated in the same pass, native multi-shot (characters/environment/lighting/sound held across cuts), RAW/EXR workflows. Duration conflict, both cited: pinggy says up to 20 seconds per pass (Extend pipeline can lengthen), innfactory says 121 frames at 24fps (~5s) - not settled. VRAM: 80GB+ full, distilled/FP8 variants from 12GB on one card, about 66GiB total download. License: free commercial use under $10M annual revenue, above that requires a Commercial Use Agreement. VFX positioning: HDR/restore/colorize/day-to-night LoRA family, AlphaGen matte extraction, Tiled Fusion 4K/8K upsampling. Ecosystem: ComfyUI (int8_convrot builds), Diffusers 0.40.0, ltx-trainer fine-tuning.

EmbeddingGemma 2: The Open Embedding Model With Fine Print

Google released the open embedding model EmbeddingGemma 2 on October 6, 2026 (HF repo google/embeddinggemma-2; API-verified: ungated, license tag apache-2.0, 21,148 downloads). Gemma 4 architecture: a 270M text/code backbone plus 170M vision and 300M audio encoders = 740M total, loading in four tiers (270M/440M/570M/740M) from one checkpoint; text, code, images, audio and video share one 768-dimensional space, with 8,192-token context (4x the first gen). On-device ledger (official): about 191MB active RAM for text-only and 567MB full multimodal on a quantized Pixel 11 Pro; weights are a 1.49GB bfloat16 safetensors. Benchmarks are all Google self-reported (no independent eval, no paper, MTEB submission unaccepted): MTEB Code 68.76 to 78.68 is the only meaningful gain; multilingual text moved just +0.21 and English slightly declined - the story is "same text quality, new modalities, much smaller deployment," not "better everywhere." Matryoshka trimming takes 768 dims down to 128 for up to 6x storage savings. License nuance, honestly: Apache 2.0 declared in metadata and ungated (v1 was gated under the Gemma license - a real opening up), but the repo root has no LICENSE file and the model card retains a Gemma Prohibited Use Policy line. Ecosystem: sentence-transformers 6.1.0+, transformers 5.19.0, Ollama, llama.cpp, MLX, Qdrant; the first-gen model passed 20 million downloads.