DFlash 2: a 3.43× drafter that fades to 1.45× under load
Inco AI's block-diffusion drafter reaches 3.43× on GSM8K with one request in flight and 1.45× at concurrency 32. Here is what you pay in memory and variance.
What the AI world is shipping this week
New open-source projects and models, ranked by attention — with our own breakdowns of the ones worth your time.
Week of Sep 15, 2026 – Sep 22, 2026
A snapshot of what got the most attention that week, not a complete index.
Latest breakdowns
All breakdowns
Jev Ultrafast: the browser agent that stopped writing text
browser-use's jev-ultrafast finishes a Google Flights search in 7.1 seconds by replacing the agent loop with one typed decision per cycle. Here is the catch.
Laya on the Neural Engine: 4.98 ms and 0.154 J per decision
mizorewww's Core ML port runs one Laya typed decision in 4.98 ms and 0.154 J on an M3 Max. What the parity fixtures prove, and what they do not.
Nimble: 2,676 examples to get within three points of Jev
Bespoke Labs fine-tuned Qwen3.5-9B on 2,676 synthetic examples. It reaches 90.12% against Jev's 93.21%, and its evaluation covers six of ten trained domains.
Trending repositories
| Rank | Project | Stars | Links |
|---|---|---|---|
| 01 | browser-usejev-ultrafast Fastest and cheapest web agent 16,948 stars | 16,948 | |
| 02 | NandhaKishorMlaya No description provided. 12,852 stars | 12,852 | |
| 03 | TheoLeeCJSemIf Semantic ifs from open models, on a 3090 at home. Independent; not affiliated with Jev or TypeSafe. 3,475 stars | 3,475 | |
| 04 | jaredpalmerkev tiny Jev-like family of decision models built on top of Qwen3.5 you can train and run on your own 2,828 stars | 2,828 | |
| 05 | mcncarljianying-headless Private source preview: native Jianying drafts, isolated editing/export, and standalone Agent Skill. 2,363 stars | 2,363 | |
| 06 | bespokelabsainimble Local typed decisions, contrastive data curation, and model evaluation. 1,549 stars | 1,549 | |
| 07 | newliver666apk-reverse Suitable for Android APK reverse engineering analysis 818 stars | 818 | |
| 08 | awlevintypesafe-computer-use Computer use for about $0.0002 a step: OCR the screen, classify the next action with TypeSafe, click. macOS. 777 stars | 777 | |
| 09 | korcarctext-humanizer text-humanizer is an open-source project designed to convert text generated by AI to its "humanized" version. Bypasses the most of AI detectors such as Turnitin or GPTZero 740 stars | 740 | |
| 10 | incoaisplash A local inference engine for Apple silicon, built around the model. 587 stars | 587 | |
| 11 | arvindearwp2shell-PoC CVE-2026-63030 & CVE-2026-60137 RCE chain proof-of-concept 563 stars | 563 | |
| 12 | ctdalcve-2026-41940-PoC A cPanel and WHM authentication bypassing tool 528 stars | 528 |
Trending models
| Rank | Project | Likes | Links |
|---|---|---|---|
| 01 | QwenQwen-Image-2.1 We are excited to open-source Qwen-Image-2. 983 likes | 983 | |
| 02 | XingChen-AGIXing4.0-29B-A4B This repository provides the model weights and configuration files for Xing4. 410 likes | 410 | |
| 03 | incoaiQwen3.8-27B-DFlash2 Blog | GitHub This repository contains the DFlash 2 draft model for Qwen/Qwen3. 236 likes | 236 | |
| 04 | espnetkan-bayashi_ljspeech_vits bash pip install espnet espnet tts "Hello from ESPnet" -o out. 226 likes | 226 | |
| 05 | fastinogliner2.5-multi-v1 *Extract entities, classify text, parse structured records, score span attributes, and extract relations — all in one boundary architecture. 222 likes | 222 | |
| 06 | XGENlabsXGEN-JING - x] JING-Flash-v1 — Four-step bidirectional model. 137 likes | 137 | |
| 07 | ISTA-DASLabQwen3.8-Flash-Next-GSQ-RCO-GGUF This repository provides GGUF quantizations of Qwen3. 129 likes | 129 | |
| 08 | soyaakinoharaqwen3.8-27b-abliterated-3.69bpw-12GB-MTP.gguf A Ridge-style mixed-quantized GGUF of the BF16 AEON-7/Qwen3. 106 likes | 106 | |
| 09 | nvidiaQwen3.8-27B-NVFP4 Description: The NVIDIA Qwen3. 104 likes | 104 | |
| 10 | yandexAliceAI-Foundation-80B-A3B-Base English version AliceAI-Foundation-80B-A3B-Base – базовая языковая модель с гибридной архитектурой и MoE-слоями. 99 likes | 99 | |
| 11 | Cactus-Computeneedle3 !Needle 3 at a glance Needle 3 is a Laddered Simple Attention Network, our small-model recipe: a Monarch Hadamard MLP in place of the FFN, GQA attention with causal conv taps, engram n-gram memory... 92 likes | 92 | |
| 12 | swiss-aiApertus-v1.5-70B No description provided. 90 likes | 90 | |
| 13 | TaoLiveAIGCTaoMate-H3 No description provided. 81 likes | 81 | |
| 14 | moondreamparakeet-redux A 1. 79 likes | 79 | |
| 15 | LiconStudioLTX-2.5-Multiple-Subject-Reference Licon MSR V1 is a multi-reference LoRA trained for LTX-2. 77 likes | 77 |
jev: 3 more projects this week
One release set off a wave of ports and clones. They are listed here instead of filling the feed.
Qwen3: 1 more projects this week
One release set off a wave of ports and clones. They are listed here instead of filling the feed.
About this data
Trend data comes from hype.replicate.dev, an open aggregator of what is gaining attention across GitHub and Hugging Face. We capture a snapshot each week and group near-duplicate projects into a single story. The aggregator caps any response at 490 items, so this is a view of the week, not a complete index of everything published.
The archive starts with 2026-W39 — earlier weeks were never captured.