中/EN

Home

Open Source

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

Frontline HotspotHardcore ReviewsOpen SourceField SOP

DeepSelect: DeepSeek's High-Performance TopK Kernel for DSA

DeepSeek's official org open-sourced DeepSelect on 2026-09-10 (snapshot: 152 stars / CUDA / created and active the same day). It is a high-performance implementation of the TopK kernel used by DeepSeek Sparse Attention (DSA) plus a companion sampler; DSA powers the V3.2, V4 and V4.1 model families, and the README claims a 2~20x speedup over native torch.topk. Version 1.0.0 and bilingual deep-dive algorithm docs landed the same day. This piece explains why TopK becomes the attention bottleneck, how RadixSelect works with its single-pass scan, randomized blocks and threshold convergence, what the expected upper bound on total processed elements actually implies, and why effective memory bandwidth - not FLOPs - is the right metric here. It then focuses on what DeepSelect deliberately does not do: it covers only two workloads, Lightning Indexer (bfloat16, topk capped at 4096) and Sampling (float32, vocab around 128K), and the README advises turning sorted_index off and setting return_value=False when values are not needed. Cold take: 152 stars means very early days; the value is ecological rather than general-purpose.

LLaDA-Image: Ant Full-Open 6B Unified Image Generation Model

Ant Group's InclusionAI open-sourced LLaDA-Image, a 6B unified image generation and editing model (208 stars / Python / created 2026-08-31, snapshot 2026-09-09). One checkpoint does both text-to-image and instruction-guided editing; both backbone and DiT are diffusion models trained in a unified framework, with image-only pre-training establishing the visual prior; the Turbo variant uses Twin-DMD distillation to cut 50 steps down to 4. It scores 53.53 (English) and 53.38 (Chinese) on Qwen-Image-Bench, a double SOTA. HuggingFace and ModelScope host Base and Turbo weights, each with an FP8 variant, and community ComfyUI support landed on 2026-09-07. Biggest caveat: the repo's license field is null with no LICENSE file - confirm terms with InclusionAI before commercial use rather than assuming Apache-2.0 or MIT.

OpenMAIC: Multi-Agent Classroom That Topped GitHub Weekly

THU-MAIC/OpenMAIC topped the GitHub weekly chart with +8,095 stars in a week (33,053 stars / 5,369 forks / TypeScript / MIT as of 2026-09-08). It turns any topic or document into a multi-agent interactive classroom: AI teachers and classmates lecture, discuss, draw on a whiteboard, and speak via TTS, generating slides, quizzes, interactive simulations and PBL activities, exportable as .pptx or interactive HTML. v1.0.0 (2026-08-27) adds a chat-first agent workbench, durable sessions, and 20 built-in skills; the stack is Next.js 16 / React 19 / LangGraph 1.1. It relicensed from AGPL-3.0 to MIT at v0.3.0 and ships a standard SKILL.md package usable from OpenClaw, Codex, WorkBuddy and more.

DeepSeek Harness: A Plugin-Everything Agent Framework

DeepSeek open-sourced its agent orchestration framework DeepSeek Harness (CLI: dsh) on GitHub under MIT, written in TypeScript and built on the Cordis runtime with an "everything-is-a-plugin" architecture that modularly assembles AI pipelines. The repo was created 2026-08-13 and passed 200k stars within ~3 weeks; it is currently 0.1.3-alpha, a developer preview with breaking changes expected (read SAFETY.md first). Launch the Web UI with `npx @deepseek-ai/dsh web` at http://127.0.0.1:3080.

Alibaba Open-Sources Qwen3.8-Flash-Next, a Qwen4 Preview

On 2026-08-26 Alibaba open-sourced Qwen3.8-Flash-Next on Hugging Face and ModelScope: a 125B MoE model with 6B activated per token, the first open-weight preview of the Qwen4 architecture. Native context is 262K, extensible to 1M via YaRN; API pricing is \$0.16/\$0.47 per million tokens (about one-twelfth of flagship Qwen3.8-Max). Benchmarks: DeepSWE 58.7, SWE-bench Pro 62.5, CoWorkBench 73.9, AndroidWorld 84.5, MathVision 95.7, and rank 7 on the open Agent Arena. It ships under the qwen-community-1.0 license (not Apache 2.0), permitting commercial use and self-hosting, but verify the terms against the model page before commercial use.

DeepSeek-V4-Flash-Vision-Exp Open-Sourced: MIT Vision Build

DeepSeek-V4-Flash-Vision-Exp is an experimental vision build of V4-Flash. Verification correction: it is not on GitHub (the api.github.com repo returns HTTP 404); the open-source home is Hugging Face (deepseek-ai/DeepSeek-V4-Flash-Vision-Exp). IThome actually reported it on 2026/8/31 19:35, not 9/1. MIT license (commercial use allowed), image input for JPEG/PNG/GIF/WebP, vision "near Claude Opus 4.8" and plain-text parity with V4-Flash — all from official API docs and media. Params ~305B and 1M context are media-reported 【Unconfirmed】. This resource gives landing guidance: self-host via the HF model card, light use via the official API, and smoke-test with real business images first. Confirmed vs unconfirmed items are split in tables.

Harbor and Terminal-Bench: Verify Vendor Scores Yourself

On 2026-09-01 Anthropic published Fable 5.1/Mythos 5.1 Terminal-Bench 4.0 scores — Fable 5.1 55.8%, Mythos 5.1 60.9%, GPT-5.6 Sol 37.3% — and harbor pushed code the same day. This article shows how to use that open-source toolchain to turn vendor scores from "cite-only" into "reproducible". GitHub API measured: harbor-framework/harbor (4872★/1705 fork/Python/Apache-2.0/created 2025-08-04/pushed 2026-09-01) is the eval framework; harbor-framework/terminal-bench (594★/pushed 2026-09-01, most active) is the task and benchmark suite; the old laude-institute/terminal-bench has 301-redirected to terminal-bench-1 (2559★ but stalled at 2026-07-11) — do not conflate them. Terminal-Bench 4.0 recalibrated quotas, removed 8 tasks and fixed 19, so scores are not comparable to earlier versions. A four-step self-built eval: start with private tasks, version the harness config, repeat and report distributions, treat cost and failure modes as first-class.

A 39k-star open-source second brain, but the license is GPL-3.0: OpenHuman in depth

tinyhumansai/openhuman (GitHub API measured 2026-09-01: 39,264 stars / 3,855 forks, Rust, GPL-3.0, created 2026-02-18, 382 open issues, not archived) is a local-first "second brain." Three architectural layers: a scored Markdown-tree memory (SQLite plus an Obsidian mirror you can hand-edit and grep), checkpointed graph orchestration (not loops — pausable and resumable across restarts), and Signal-protocol end-to-end encrypted agent-to-agent comms ("No server ever sees plaintext"). TokenJuice compresses tool output ("up to 80%" is a self-reported ceiling, not a measured average). Privacy Mode is enforced in the Rust core so inference never leaves the machine. Three caveats: Early Beta plus 382 issues means no production SLA yet; GPL-3.0 is strong copyleft, get legal review before embedding in closed source; the README's four-way comparison table is stale (it still labels OpenClaw "Plugin-reliant / Single loop / no Auto-fetch" while 2.0 introduced Shared Cloud Sessions and Active Memory), and the 90k Skills / 5k MCP / 80% compression figures are all vendor-reported, unverified by any third party.

Microduck: Hugging Face's $399 Open-Source Duck Robot Ships a Full RL Stack, With Hardware Files in Dispute

On 2026-08-27 Hugging Face and Pollen Robotics launched Microduck, a 25cm duck-shaped biped weighing under 800g at $399, with preorders opening the same day. The hardware is an RK3566 with 1GB of RAM and 32GB of storage, 15 XL330 servos, an 8x8 ToF LiDAR, and a 2600mAh battery rated at about an hour, running its control policy as ONNX on board at 50Hz, which is the spec that separates it from an app-driven toy. The valuable part is the software released alongside it: runtime and SDK, physics simulation, and a complete reinforcement learning training stack, published across `pollen-robotics/microduck` and `pollen-robotics/microduck_rl`, both Apache-2.0, measured via the GitHub API on 2026-08-31 at 4108 stars / 485 forks and 1037 stars / 179 forks. Training uses MuJoCo Warp (mjlab) with PPO, producing a deployable gait in one to two hours across 4096 parallel environments, with an actuator model, domain randomisation, and backlash simulation bridging sim-to-real, and observation normalisation baked into the exported ONNX. One honest caveat: whether the hardware design files were released too is disputed across sources, so this piece devotes a section to separating the four layers rather than picking a side.

Architecture Innovation Needs Someone to Write the Kernels: Qwen Open-Sources FlashQLA, 2-3x Faster GDN Forward, Hopper and Up Only

QwenLM/FlashQLA (GitHub API, checked 2026-08-30: 670 stars / 69 forks, Python, MIT, created 2026-04-24, last push 2026-08-26) is a high-performance linear attention kernel library built on TileLang. The authors report 2-3x forward and 2x backward speedups for GDN Chunked Prefill over the FLA Triton kernel on NVIDIA Hopper and Blackwell, with the largest gains in pretraining and edge-side agentic inference. Three things worth studying: gate-driven automatic intra-card context parallelism (exploiting the GDN gate's exponential decay under TP, long sequences and small head counts), hardware-friendly algebraic reformulation (cutting Tensor Core, CUDA Core and SFU overhead without losing precision), and TileLang fused warp-specialized kernels (hand-written warpgroup specialization that balances context-parallel and backward needs). Since v0.1.2 it also serves as a GDN backend for flash-linear-attention, plug-and-play through the standard FLA API. The timing is the interesting part: it shipped the same day as Qwen3.8-Flash-Next, meaning Alibaba handed in the architecture paper and the kernel implementation together. The entry ticket is not cheap though - SM90 or newer, CUDA 12.8+, PyTorch 2.8+, and the 2-3x figure is self-reported rather than independently reproduced.