Home

Open Source

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

Qwen-Image 2.1 Open Weights Are Back, But Apache Is Not

Alibaba's Qwen team open-sourced Qwen-Image-2.1 on September 20, 2026: text-to-image, image editing and native RGBA transparency in one 7B checkpoint - a 32-layer single-stream DiT (mixed-granularity attention plus prefix KV reuse) + a Qwen3-VL 8B text encoder (which also encodes reference images) + a 64-channel RGBA VAE. About 33GB in BF16, compressible to roughly 14GB via Comfy-Org quantized repacks. Capabilities: native 2048x2048 (seven ratios up to 2752x1536), 40-step inference, up to 10 reference images per call, circle/mask local edits, direct transparent-PNG output (official prompt phrasing provided), plus two 9B prompt rewriters. Reported throughput: vLLM-Omni at 1024 squared, 40 steps, BF16 peaks at 34GB and 3.3-4.5s per image (single source); Alibaba says a 3090 runs it with offload. Day-zero ecosystem: Diffusers, ComfyUI, vLLM-Omni, SGLang, LightX2V; 14 fine-tunes and 18 quantizations within the first day. The core of this piece: every prior Qwen-Image shipped Apache 2.0 - 2.1 switches to the Qwen Research License (non-commercial only; commercial use needs a separate agreement; fine-tunes inherit the restriction and must display "Built with Qwen"). The community "license trap" pushback and the "License renders this model useless" thread are reported as-is. Qwen-Image-Bench 60.28 (seventh) is vendor-designed and vendor-run. Lineage section: 3.0 is the API-only fork, 2.1 is the downloadable line - the weights came back, the Apache license did not.

Bilibili's 35B translation MoE: 150 languages, only 3B active

Bilibili's Index LLM team open-sourced the Index-Translate family on September 30, 2026 (weights on Hugging Face and ModelScope, Apache-2.0): 2B/9B/35B-A3B-preview text models built on Qwen3.5, targeting 150 languages (self-reported), with the 35B MoE activating only about 3B parameters per token and a 262,144-token context. Three branches split the work: Index-Echo produces dubbed speech that preserves the source speaker's voice (Chinese to/from English, Spanish, Japanese); Index-Homura controls syllable counts (the 9B lands within +-10% of the target on 81.92% of SandGlass cases - built for dubbing timing); Index-NativeLong (aka Index-Nailong) keeps long documents consistent (a 32K-token fantasy text keeps a name/royal-title pun intact, where chunk-by-chunk translation drifts). Constrained translation via instTrans: hard constraints enforce glossaries (e.g. carbon fiber) plus JSON/CSV/code/placeholder structure; soft constraints cover tone, domain disambiguation (plant to factory), cross-sentence consistency and LaTeX preservation. Local deployment is one line with the official GGUF (Q4_K_M, 21.71GB) via llama.cpp serve, plus a browser extension and a video dubbing pipeline. Honesty note: every score is self-reported with no independent replication yet (FLORES COMET-22 0.8794, WMT26 Judge 76.76, 2.4% off-target on low-resource pairs), and in the same official table DeepSeek-V4.1-Flash scores higher on WMT26 Judge at 83.55 - not first place on raw quality.

NVIDIA Locks Down AI Agents: Open Runtime, Silicon Watchdog

A deep dive into NVIDIA's Open Agent Safety Platform (official materials plus multi-source reporting, October 5, 2026 basis). Core thesis: agent safety cannot rest on the model behaving itself - when the agent is itself the software trying to escape its guardrails, application-layer protections fail, and agent drift (gradual divergence from operator intent via vague prompts, missing tools or policy conflicts) triggers no alarm. Two layers: OpenShell (Apache 2.0 open source, a kernel-level sandboxed secure runtime whose policies cover files, processes, credentials, tools, network and databases; intent-alignment checks before execution plus continuous drift monitoring; runs on Vera CPUs, extensible to Arm and Intel) and NVIDIA Sentry (an out-of-band watchdog on BlueField-4 DPUs that verifies agent identity via DOCA, produces attested telemetry, enforces granular policy, and quarantines an escaping agent within milliseconds; a reference design, not open hardware). Three layers stay separate: application, runtime, infrastructure. Over 100 organizations are on board (Anthropic integrating it with Claude Managed Agents, Salesforce for Slack, SpaceXAI for Cursor agents) under the Linux Foundation's Open Secure AI Alliance. The criticism is reported faithfully: Gartner notes OpenAI, Amazon and Google are absent; IDC estimates it addresses under 25% of enterprise agentic security problems; Control Risks notes it only governs known agents on infrastructure you own (shadow agents, SaaS-embedded ones and attacker-delivered ones are out of reach); lock-in risk discussed. Together with open-weights Kimi K3 it marks the two ends of the agent-safety map.

LTX-2 Open-Sourced: One 22B Model Generates Video and Sound

A deep dive into LTX-2 as open weights (GitHub verified 2026-10-02): Lightricks/LTX-2 at 9,569 stars, Python, last push October 2; per the official README, the first DiT-based audio-video foundation model that generates picture and synchronized sound in one pass. LTX-2.5 composition: a 22B distilled transformer (bf16) plus a custom Gemma 4 12B text encoder (not interchangeable with Google stock), video and audio VAEs, spatial and temporal upscalers - roughly 66 GiB in total; DistilledPipeline runs on just 8 preset sigmas (8-step stage 1 + 4-step stage 2) for the fastest path, FP8 quantization and CPU/disk offload cut memory, and 4K means 3840x2176 (the README explicitly says not 2160). Twelve pipelines include DFR production quality, DubIt re-dubbing with lip-sync preserved, Retake partial regeneration, and SDR-to-HDR (BT.2020/HLG plus ACEScct EXR); ltx-trainer covers LoRA, full fine-tuning and IC-LoRA, with an official ComfyUI plugin. License verification: the GitHub badge reads NOASSERTION because LTX-2 ships a custom LTX Community License (applying to LTX-2.5 since August 11, 2026), not an OSI-approved open-source license - free for personal non-commercial use, but entities with annual revenue of USD 10 million or more must purchase a Commercial Use Agreement for any commercial use; derivatives include distillation, and redistribution must carry the full agreement. The self-hosting case: data stays in-house, marginal cost of batch generation approaches electricity, and LoRA styles remain your own asset.

Open-source Dots: 2,400 stars in three days

A teardown of feder-cr/dots (GitHub API snapshot 2026-10-02): 2,416 stars, 421 forks, MIT-licensed Python, created 2026-09-29 - an open-source isotope that hit escape velocity the day after OpenAI's closed Dots debut. Its core thesis: the model is swappable with one flag, the browser is what the website actually sees. A real Firefox engine patched in C++ decides the fingerprint inside the engine rather than as a page-inspectable JavaScript coat; one seed equals one consistent identity (screen, fonts, GPU, timezone and language agree, reproducibly); nothing for a page to find (no WebDriver flag, no DevTools protocol, no automation globals); the pointer travels before it clicks and keys are pressed one at a time so events arrive trusted; --profile-dir keeps logins across runs; with --proxy the timezone and language follow the exit. Models come from OpenRouter with a one-flag swap, and the companion repo invisible_playwright_mcp exposes the same browser as an MCP server for Claude Code, Codex and Gemini CLI. The README states plainly it is not affiliated with OpenAI. Includes a compliance boundary note: respect target-site terms and local law; no fraud, ticket-scalping or bulk sign-ups.

410 TFlops, 95 Percent of Peak: DeepSeek Opens Its Ascend Stack

A repo-by-repo teardown of DeepSeek's Ascend kernel stack (2026-10-01 GitHub API snapshot): TileKernels, 1883 stars, MIT (dozens of TileLang-DSL operators - MoE routing, FP8/FP4 quantization, Engram gating, manifold hyper-connections - with an Ascend backend added 09-30 that auto-selects at runtime, acknowledging Huawei's engineering support); DeepGEMM-Ascend, 418 stars, MIT (fully API-compatible with the main repo, Ascend 950 first release, CANN 9.20 + torch_npu + Python 3.10+, Ascend scaling-factor packing differs from NVIDIA); DeepEP-Ascend, 190 stars (EPBuffer V2.5-compatible, validated on 950DT + CANN 9.2.0 + torch_npu 2.13.0rc1, and - as of the snapshot - no LICENSE file in the repo root, so commercial use awaits clarification); FlashMLA, 13015 stars, MIT (Ascend sparse attention at 410 TFlops / 95% peak prefill and 360 TFlops / 83% decoding, versus 1460/950 TFlops on B200; fused kernel is CUDA-only). Cross-referenced against the NVIDIA-side repos (DeepGEMM 7,907 stars, DeepEP 10,231); treat each repo's LICENSE file as the final license authority.

Youdao Open-Sources the Real-Time Voice Trio, All Apache-2.0

NetEase Youdao open-sourced its Ziyue Live real-time voice trio (2026-09-29 GitHub API snapshot: Confucius4-R2T2 at 680 stars, Confucius4-T3PO at 110, Confucius4-TTS at 821; all three LICENSE files are verbatim standard Apache-2.0 text, verified word-by-word). R2T2 is true-streaming ASR: append-only output that never revises committed text, 200-600ms average latency, configurable 80ms-to-2s decode chunks, built on Qwen3-ASR with a Longest-Stable-Prefix paradigm, shipping vLLM/transformers/llama.cpp backends plus a WebSocket server. T3PO is a 14B text-to-text simultaneous-translation model with streaming READ/WRITE decisions, interleaved-history KV-cache reuse, and Pareto-aware RL jointly optimizing quality and latency; cascade it with R2T2 for a speech-to-translation pipeline. TTS covers zero-shot cloning across 14 languages with cross-lingual voice preservation (arXiv 2608.11650). R2T2 and T3PO topped the HF ASR and Translation Trending boards (official press basis); weights are open on both HF and ModelScope.

Xiaomi Open-Sourced the RL Stack That Trained MiMo-V2.6

A fact-check of XiaomiMiMo/verl (2026-09-27 GitHub API snapshot: 465 stars, 49 forks, Python, Apache-2.0, created 2026-09-21, pushed 2026-09-26). It is the Agentic RL training code behind MiMo-V2.6, forked from verl-project/verl (23,650 stars, HybridFlow) and built on verl 0.9.0.dev with reproduction code for five RL environment suites: Code software engineering (executable tests), Cyber vulnerability reproduction (rule checks), General knowledge work (rubric review), Visual web development (visual scoring), and Music symbolic composition (rule checks), each with launch scripts. The MiMo-V2.6-RL-oss training dataset on Hugging Face and the technical-report PDF are open too; same-day companions uni-agent (14 stars) and mimoagent (30 stars, MIT) complete the bundle. Divergence from upstream cannot be verified, so the article sticks to README-verifiable facts; treat the repository LICENSE file as the final license authority.

Step-Code Open Source: StepFun's Terminal Agent at 438 Stars

A fact-check of stepfun-ai/Step-Code (2026-09-27 GitHub API snapshot: 438 stars, 41 forks, TypeScript, license field MIT, created 2026-06-01, pushed 2026-09-26, 44 open issues). A terminal coding agent covering the full task loop of reading code, editing code and running tests; it binds to the Step provider (auto-discovers Step models at login, the opposite of MiniMax Code's BYOK multi-provider route); it supports MCP servers, Agent Skills and multi-agent orchestration out of the box; the /goal command delegates long-horizon tasks; StepPage publishes a local page as a live static site in one command. Vendor-reported figures: Terminal Bench 2.1 pass rate 80.9% tied for first, self-built Multi-Frame benchmark 73.3% on top, lowest average 5.09M tokens — all tagged vendor self-reported and not independently retested.

Ant Ming-Image: MIT is free for commercial use, Qwen is not

A fact-check of inclusionAI/Ming-Image (2026-09-24 GitHub API snapshot: 91 stars, 6 forks, Python, MIT). One design line ships two weights: Design for text-to-design, producing UI, dashboards, infographics and posters end to end, and Design-Layer, which splits a flat design into 2 to 9 semantically independent RGBA transparent layers. The architecture is a 6.15B design transformer plus a 17.01B multimodal LLM plus a 3.09B connector, about 26.44B total (vendor basis). It offers a native RGBA VAE, 8K structured prompts, and sampling defaults of steps 12 and CFG 1.0 for Design (2.0 for Layer), with recommended resolutions 2048 and 1024. The license red line: MIT allows commercial use, self-hosting and derivative work, contrasting with Qwen-Image 2.1's non-commercial Qwen Research License. The hardware floor is a single 80 GiB GPU at BF16; speed and benchmark figures are tagged vendor basis and not independently retested.