Every week, I organize 400 to 700 links into roughly 60 categories as part of my ongoing effort to learn about AI. This is my personal notebook, which I enjoy sharing with friends… a hobby and a labor of love, rather than a commercial publication or product.
If you arrived here through a search or shared link, this page collects the links I found for International for the week ending September 4, 2026.
As part of my learning process, I like to automate the category covers. It gives me a chance to learn Python and APIs.
This week’s cover prompt was written using Claude Opus 4.7, and the image was generated using Nano Banana 2.1.
Category Cover Image Prompt: Vertical psychedelic screen-print gig poster of a central illustrated globe with continents melting and blooming into swirling op-art patterns, orbiting rings of dotted tour-route lines connecting glowing cities across every continent with tiny paper airplanes and autumn leaves drifting along the paths, bold hand-lettered title INTERNATIONAL arcing across the top in 1960s Fillmore-revival psychedelic letters, electric magenta acid orange deep teal and lime green ink on cream paper with visible registration offset and ornate decorative border.
This Week in International News
Summary by Claude Sonnet 5.5 based on this week’s links:
- Chinese open models dominate: GLM-5.3 opened its weights (744B total, 40B active, 1M context), and Unsloth says a 2-bit version runs on a 256GB Mac. Tencent's open Hy4 Preview (770B) also landed near the top of the Code Arena WebDev leaderboard.
- Alibaba's Qwen and Wan push the leaderboards: Arena reported Qwen3.8-Max-0902 debuting at #1 in WebDev with 1691 points. Wan 3.0 took #1 in the Video Edit Arena.
- Open-sourced agent harnesses: A Zhihu contributor argues that DeepSeek and OpenAI open-sourcing their harnesses makes the runtime a possible next platform. Separately, ByteDance Seed's HarnessDev paper evaluates the harness an agent builds, not just its task results, and finds generated harnesses lag human-engineered ones.
This summary was generated by Claude Sonnet 5.5 to help you explore the links below. Rest assured, I select, organize, and check the links by hand in Google Sheets, and write the introduction and personal commentary in The Main Newsletters myself each week as a labor of love.
This week's links related to International
Banger paper from ByteDance Seed. If you are curious about self-evolving agent harnesses, this one is worth your time. (bookmark it) The proposes method, HarnessDev, stops scoring a model on the tasks it completes and scores it on the harness it builds. The agent starts from
https://x.com/omarsar0/status/2095170896407548190
🧩 DeepSeek and OpenAI Open-Sourced Their Harnesses. The Runtime May Become the Next AI Platform DeepSeek Harness and OpenAI’s Codex harness are now open source. But the larger story is not simply that two more repositories became public. Zhihu contributor 第欧根尼 argues that
https://x.com/ZhihuFrontier/status/2093253880482316422
DeepSeek Harness Did Not Stop Updating. It Just Shipped a Breaking Rewrite. The premise is already outdated. DeepSeek Harness released v0.1.2-alpha.1 (https://t.co/fC0UNEn6LB) on August 27, followed by v0.1.2-alpha.2 (https://t.co/DhdSUmqnn9) on August 30. The apparent silence
https://x.com/ZhihuFrontier/status/2094348274291691531
42 models are now available through the Perplexity Agent API, including the open-weight GLM 5.3 from @Zai_org. Switch models by changing one field in your request. Add fallbacks to keep your agent running if a model becomes unavailable.
https://x.com/perplexitydevs/status/2094945628426256638
Amazon and Meta AI bots account for most website visits – Press Gazette
https://pressgazette.co.uk/platforms/amazon-and-meta-ai-bots-account-for-most-website-visits/
🎬Video generation faster than playback! 🚀MiniMax H3 on vLLM-Omni + FastVideo’s FastH3: a complete 10.1s MP4 – video AND synchronized audio – rendered in 8.7s!⚡️ Thanks to @MiniMax_AI for the great Minimax H3 release, the FastVideo team @haoailab for open-sourcing FastH3 and
https://x.com/vllm_project/status/2094849929487552663
How many cuts can a video model handle in 15 seconds? MiniMax H3 Max is pushing the limits.
https://x.com/fal/status/2093147720898736495
Minimax H3 Max has generates video faster than you can watch it so I hooked it to a twitch livestream! Now you can watch infinite interdimensional cable – link to the stream below
https://x.com/rehan_shei/status/2093528415576211819
Scoop: The US start-up Together AI is partnering with Humain, the Saudi Arabian AI company, on a 250 megawatt data center … where it can bypass U.S. backlash over data centers. The Saudi-based data center will be among the largest publicly disclosed data centers built
https://x.com/nikogallogly/status/2094394048844894487
We just signed one of the largest AI infrastructure deals for open source, period. 250MW data center. $5B+ in annualized revenue. Built with @HUMAIN in Saudi Arabia.
https://x.com/togethercompute/status/2094416469920796999
Manus Resumes Independent Operations
https://manus.im/blog/manus-resumes-independent-operations
GLM-5.3 can now be run locally! The 2-bit model retains ~81% accuracy after we shrunk it from 1.51TB to 239GB (-83% size). Run on a 256GB Mac or RAM/VRAM setups. GLM-5.3 is the strongest open model to date. Guide:
https://t.co/NLgb3CMB6A GGUF:
https://x.com/UnslothAI/status/2093397494889890050
🎉 Congrats to @Zai_org on opening the GLM-5.3 weights, the largest model in the GLM-5.3 line. Day-0 support in vLLM. 744B total, 40B active, 1M context, 128K max output. @Zai_org kept the GLM-5.2 base and scaled post-training instead, so nothing under the model changed. vLLM
https://x.com/vllm_project/status/2093354756244992383
Today, we’re releasing GLM-5.3 Fast: one of the most intelligent open-weight models ever at an even higher TPS. Designed for real-time use cases that demand consistent performance. Get access here:
https://x.com/baseten/status/2095338689492578693
Qwen3.8-Flash-Next by @Alibaba_Qwen has landed in Agent Arena, ranking #7 among open models (#24 overall) with +2.4% net improvement across 8.7K+ real-world agentic sessions! Among open models, it sits just behind DeepSeek V4 Flash (High) at #6 (+3% net improvement) and two
https://x.com/arena/status/2094566204488962483
🏆 #1 on CodeArena: WebDev leaderboard. Qwen3.8-Max-0902 jumps from 1669 to 1691, setting a new record for agentic coding (WebDev) workflows, with standout strength in multistep reasoning, tool use, and full app generation. Thanks for the recognition! @arena
https://x.com/Alibaba_Qwen/status/2094976556494209206
Banger paper from the Qwen team. If you evaluate agents on anything longer than a single session, this one is worth your time. (bookmark it) E-Commerce Bench runs an agent through a simulated 365-day year operating several online stores at once. 18 frontier models are scored
https://x.com/dair_ai/status/2094872928240447665
Hermes now has a local backend! There’s support for Unsloth UD-Q4_K_XL and UD-Q4_K_M quants for DeepSeek-V4-Flash, Qwen3.8-27B, 3.6-35B-A3B and Qwen3.8-Flash-Next in Hermes!
https://x.com/danielhanchen/status/2095623899979600152
If you build agent skills in production, check out this great paper from Alibaba. You can think of a production agent skill as a directory. The root loads on activation, and references, schemas, scripts, assets and nested subskills load only when an execution path reaches them.
https://x.com/dair_ai/status/2094811526767182090
GLM-5.3 at 310 tok/s! Databricks inference is #1 in both speed and latency, again. On our internal Databricks coding benchmark, GLM-5.3 is the strongest OSS model for coding, competitive with Fable 5 and Opus 4.8.
https://x.com/Yuchenj_UW/status/2094993931268420072
Also, Anthropic made cache read prices 75% cheaper. This is only 13x more expensive than deepseek flash now for cache reads (most of the tokens paid for)
https://x.com/Teknium/status/2094861678785806595
Can LLMs build and evolve their own agent harness? ByteDance Seed introduces HarnessDev, a benchmark that shifts evaluation from task outputs to the runnable infrastructure itself. Across six creator LLMs and five benchmarks, generated harnesses lag human-engineered ones on code
https://x.com/HuggingPapers/status/2095545764793520204
Exciting news: GLM-5.3-Flash by @Zai_org has landed in Agent Arena! At a $0.12 median cost per task and a +4.6% net improvement, it has reshaped the Pareto frontier! GLM-5.3-Flash is placed between DeepSeek V4 (High) and GPT-5.6 Luna (xHigh). Based on 9K+ real-world agentic
https://x.com/arena/status/2094440382440611935
Banger paper from BAAI. If you are building research agents, this one is worth your time. (bookmark it) They find that adding skills scores 134.3% higher on MLE-bench, 34.4% higher on PaperBench, 9.2% higher on FrontierCS and 14.0% higher on PassNet. More details on the
https://x.com/dair_ai/status/2095539831141220620
DeepSeek-V4-Pro-0813 is live on CoreWeave Serverless Inference. 1.6T parameters. 1M context. Built for long-horizon agent work, and priced for it: cache reads run $0.044/M for all the context your agent re-sends every step. Start here:
https://x.com/CoreWeave/status/2094878660217995750
Kimi Code New Feature: Remote Control Kimi Code 0.39.0 add Remote Control as an experimental feature for accessing a local web session remotely. Enable it with KIMI_CODE_EXPERIMENTAL_REMOTE_CONTROL=1, then run kimi rc, kimi web –remote-control , or /remote-control to start
https://x.com/KimiDevs/status/2093184808419746164
Tencent just released Hy4 Preview: a 770B-parameter open-source model with 49B active parameters and a 1M-token context window. The most interesting part is its agentic research capability. Tencent says Hy4 coordinated several Codex sessions in parallel, evaluated their results
https://x.com/kimmonismus/status/2093237109708468361
Tencent Hy4 Preview Has Entered China’s Top Agent Tier @TencentHunyuan has just released and open-sourced Hy4 Preview. The MoE model has 770B total parameters, 49B active parameters, and a context window exceeding 1M tokens. It is positioned around coding, office work, research,
https://x.com/ZhihuFrontier/status/2094345125203992756
GLM-5.3 is now open-weight. Our most capable model for agentic coding and cyber defense is now available to download, run, and customize. Weights:
https://t.co/v1IbWMXxg4 Tech blog:
https://x.com/Zai_org/status/2093354097122455713
Yesterday, we rolled out a configuration update for GLM-5.3-Flash to improve performance in some agentic use cases. If it underperformed Ox Alpha in your workflow between August 26 and 27, please give it another try and let us know if the experience has improved.
https://x.com/ZixuanLi_/status/2093328501520663007
AIエージェントに身体を与える卓上型AIロボット「Palmimo DevKit」、本日より先行販売を開始します！ 目・耳・口・顔・脚、そして知性。数行のPythonで動き、頭脳のAIは差し替え自由。ロボットやハードウェアの知見がなくても、自分のアプリケーションでロボットを動かせます。
https://x.com/maze_rapid/status/2095294835364364337
Interesting paper from Tencent. Tencent trains an agent to manage its own working context, and assigns credit at the level of individual context edits. Long-horizon tasks force a model to retrieve, integrate and maintain scattered information across many turns, and keeping
https://x.com/omarsar0/status/2094505508850032852
Big news: Qwen3.8-Max-0902 by @Alibaba_Qwen just debuted at #1 overall in the Code Arena: WebDev with 1691 pts! It scores 3 pts above Claude Opus 5 (Max), 17 pts above Kimi K3 (Max), and 22 pts above the previous Qwen3.8-Max. Priced at a blended $5/MToken, Qwen3.8-Max-0902 also
https://x.com/arena/status/2094974637704913198
Wan 3.0 debuts at #1 on the Artificial Analysis Video Editing Leaderboard, and is a close #2 in Text to Video with Audio Wan 3.0 is Alibaba’s new all-in-one video generation and editing model, positioned as a single system for turning multimodal creative direction into video. It
https://x.com/ArtificialAnlys/status/2095349174799888760
Exciting news: Wan 3.0 by @Alibaba_Wan has claimed the #1 spot in the Video Edit Arena! With 1414 ps, Wan 3.0 leads dreamina-seedance-2.5 by 4 pts and MiniMax-H3 by 22 pts. This release marks @Alibaba_Wan’s debut in the Video Edit Arena. This newer Arena now features 10 models
https://x.com/arena/status/2093143153167810608
Qwen3.8-Flash-Next is #24 overall with +2.4% net improvement (#7 among open), and by signal: – Confirmed Success: +12.3% – Bash Recovery: +2.9% – Praise vs. Complaint: -1.6% – Steerability: -1.1% – No issues with Tool Hallucination
https://x.com/arena/status/2094566207794061800
🚀Qwen3.8-Max just got upgraded. Meet Qwen3.8-Max-0902! 2.4T parameters. 1M context tokens. Built for real world complexity. Further post trained on Coding & Cowork, Qwen3.8-Max-0902 now delivers stronger performance across complex enterprise tasks, scientific research, and
https://x.com/Alibaba_Qwen/status/2094968708288680276
LESSON LEARNED: Always use BF16 kv cache. I was using turboquant. Yeah the VRAM consumption sucks – so another trick is to CPU Offload the kv cache. Implemented in SlimServe for 3090 and Qwen (it requires extensive testing to get it right) I have further work to universalize
https://x.com/QuixiAI/status/2093405502181179422
125B/6B, 1M context, multimodal, now in OpenCode Go. ⚡ Happy coding with Qwen3.8-Flash!
https://x.com/Alibaba_Qwen/status/2093227357951897687
My experience with Qwen3.8-Flash-Next is that it can’t even keep track of a multi-turn conversation. It answers questions I asked 2 turns ago. This is at FP8
https://x.com/QuixiAI/status/2093175458569326919
Qwen3.8 Flash is the third Qwen3.8 release, after Max and 27B it uses a 125B MoE architecture with only ~6B parameters active per token Flash is ~20x cheaper and ~2x faster than Qwen3.8 Max, often delivering performance close to Max hosted version comes with a 1M token context
https://x.com/skalskip92/status/2093384847649571325
Today, Transluce is releasing the most expansive independent evaluation to date of how AI systems respond to users experiencing mental health crises. We evaluated 77 model variants from OpenAI, Anthropic, Google, Meta, SpaceXAI, Thinking Machines, DeepSeek and Moonshot AI.
https://x.com/TransluceAI/status/2094455208759693476
GLM-5.3 is a good model, and as the open weights models get better and better it becomes increasingly important that they actually publish model cards, do red teaming, etc. Since you can break the guardrails with any open model, we need a sense of what the risks are as well.
https://x.com/emollick/status/2093399712208494741
LongCat-2.0 is free in Cline right now. It’s a 1.6T open weights MoE model with 1M context from @Meituan_LongCat scoring similar to Claude Opus 4.7 and Gemini 3.1 Pro. Try now: npm i -g cline /model Choose LongCat-2.0 under free models
https://x.com/cline/status/2094903089409261667
We are rolling out GLM-5.3 on Ollama. Private. Fast. US and Europe hosted. No data retention. Try Ollama’s GLM-5.3-Flash as we bring it GLM-5.3 online. Super fast. Claude Code: ollama launch claude –model glm-5.3-flash:cloud OpenCode: ollama launch opencode –model
https://x.com/ollama/status/2093356025084797176
How soon until an AI agent can organically get a million followers on a video platform like youtube or tiktok – and do so fully autonomously in a couple of months?
https://x.com/bilawalsidhu/status/2094081526459716053
Imagine when this type of just-in-time media generation is connected to something as powerful as tiktok’s for you algorithm. The feed could almost read your mind and materialize videos right before you scroll to them. It would be the most addictive form of media ever created.
https://x.com/bilawalsidhu/status/2093856257165668467
kausable is building AI systems that can adapt when the world changes. Next week I have @JohannesHaux, co-founder and CEO of kausable, on the podcast. They are building in Heidelberg, right in the wider @Cyber_Valley ecosystem. And with Cyber Valley turning 10 on September 9,
https://x.com/IlirAliu_/status/2093972489823269073
MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities – MiniMax Research | MiniMax
https://www.minimax.io/blog/minimax-h3
UK AISI Evaluation of GPT-6-Astra – No-CoT math time horizon: “Astra can solve significantly more difficult math problems in a single forward pass than past models. UK AISI measured Astra’s time horizon at 30.9 minutes compared to 3.6 minutes for GPT 5.6 Sol” – CoT
https://x.com/AiBattle_/status/2095598057857614053
WOW OpenAI: “no-CoT time horizon may have increased by about an order of magnitude” “UK AISI measured Astra’s time horizon at 30.9 minutes compared to 3.6 minutes for GPT 5.6 Sol”
https://x.com/scaling01/status/2095597192035664348
Tencent Hy4 Preview leads on SWE-bench Pro. 770B, 49B active, 1M context, and their biggest generational leap measured to date. It’s exciting to see another open weights model compete against the frontier. Try in Cline with: 1. npm i -g cline 2. /model 3. Select Hy4 preview
https://x.com/cline/status/2093401313203892241
GLM-5.3-Flash at 270 tok/s! Databricks inference team is on fire. So proud of the team! On our OfficeQA Pro v2 benchmark, GLM-5.3-Flash delivers 10% higher quality than GLM-5.2 at just 1/10 the cost. Fast, cheap, and good.
https://x.com/Yuchenj_UW/status/2093177892356472978
GLM-5.3-Flash is #19 overall with 4.6% net improvement (#4 among open), and by signal: – Confirmed Success: +15.3% – Praise vs. Complaint: +4.9% – Steerability: +2.4% – Bash Recovery: -0.7% – No issues with Tool Hallucination
https://x.com/arena/status/2094440384592298478
Big news: Hy4 preview by @TencentHunyuan just landed ~#5 in the Code Arena: WebDev with 1633 pts (AutoEval). This is a significant improvement from Hy3 at #31 overall (+115 pts)! Among open models, Hy4 preview is ~#3, compared to Hy3 at #7. Note: this is an early AutoEval
https://x.com/arena/status/2093224696745492802
Reference-to-Video is now live for MiniMax H3 Max on fal! This is an early preview that can achieve up to RTF=1 (and with references around RTF=1.5) on 768p, and we aim to improve the speed by 2x while slashing costs during this week. Quality wise it delivers the same
https://x.com/fal/status/2094527664040124764#m
There’s a fascinating new startup using dogs to detect cancer in human breath In a new trial, the dogs detected it over 90% of the time! How it works: Dognosis has patients breathe into a cotton surgical mask, which is then frozen and shipped to their farm in India. There, a
https://x.com/rowancheung/status/2094451656243265958
nvidia/DeepSeek-V4-Pro-0813-NVFP4 · Hugging Face
https://huggingface.co/nvidia/DeepSeek-V4-Pro-0813-NVFP4
This is what open-source intelligence is for. Built on vLLM-Omni and @haoailab’s FastH3, with @NVIDIAAI hardware support – all public. Real-time generation makes interactive video possible; an open baseline makes it improvable for everyone. Take it and go faster! 🚀
https://x.com/MiniMax_AI/status/2094926136333787512
Across Europe, AI communities have been moving the field forward for years. This October, Berlin becomes a meeting point for the people building what comes next. 🤝 Join researchers, developers, engineers, and industry leaders at #NVIDIAGTC Berlin, October 20-22. Register
https://x.com/nvidia/status/2095187846462054455
@TencentHunyuan’s Hy4-preview runs in vLLM from day 0, verified on NVIDIA GPUs. 🎉 – 770B total, 49B active, 256 routed experts plus one shared – 1M context, but each query attends to just 2048 tokens – Only 21 of the 78 layers compute their own sparse index, the other 57 reuse
https://x.com/vllm_project/status/2093248073057357905
Helping keep race week focused on what matters, @AstonMartinF1
https://x.com/cohere/status/2095572785288335527
I think we’ve all been sleeping on Waterloo what a city
https://x.com/cohere/status/2093781127051784535
DeepSeek-V4-Flash-Vision-Exp weights are released. Now there’s some feature parity with Moonshot and GLM.
https://x.com/teortaxesTex/status/2094375909868368213
The Arcee API now includes GLM-5.3 and GLM-5.3-Flash. Two weeks ago we opened the catalog beyond Trinity so you can run your favorite open models from one endpoint: Trinity, DeepSeek, Kimi, Inkling, and now GLM. ➡️ Use them in nac for long-running, hands-off work.
https://x.com/arcee_ai/status/2094964589775479266
UK AISI says that “Astra performed a range of malicious actions including conducting supply chain attacks against open source providers”
https://x.com/scaling01/status/2095596612856741902
The hardest part of scaling autonomous driving is not just collecting more data. It is learning from the rare, safety-critical scenarios that are easy to miss inside massive datasets. At @CVPR, Mobileye CTO Prof. @shai_s_shwartz introduced Meteor, Mobileye’s approach to
https://x.com/Mobileye/status/2095512568227865035
GLM-5.3 is officially live on Hugging Face. What should run it locally: – FP8: 10-12× H100 or 8× H200 – 4-bit/NVFP4, roughly 390-430GB: one 512GB Mac Studio or 4× DGX Spark – Aggressive 2-bit, roughly 230-250GB: one 256GB Mac Studio or 2× DGX Spark, with quality and context
https://x.com/kimmonismus/status/2093354978534477956
🚀 Hy4 preview is here. 770B, 49B active, 1M context. Built for productivity. Open source frontier. Consistent affordable price. Use it. Tell us what breaks. More on Hy blog：https://t.co/rbl1IWRk3C HuggingFace：https://t.co/mE9wevH5XR Github：
https://x.com/TencentHunyuan/status/2093222928720761009
tencent/ContextPilot-14B · Hugging Face
https://huggingface.co/tencent/ContextPilot-14B
Introducing Quasar 438B: Europe’s Leading AI Model
https://multiversecomputing.com/resources/introducing-quasar-438b-europe-s-leading-ai-model
Public attitudes toward AI will shape how governments regulate it, how firms invest, and how quickly new tools spread. Stanford researchers Michael Tomz and Diyi Yang are conducting a systematic comparison of public opinion on AI in the US and China … one of three projects
https://x.com/StanfordHAI/status/2093490181643153520
Tencent Hy
https://hy.tencent.ai/research/hy4-preview
London, your ride has arrived. 🇬🇧 Today, Wayve and @Uber launch supervised autonomous rides across all of London (except airports) – the first time the public can experience the Wayve AI Driver and the first autonomous trips available in the UK. Taking your first ride? Tag us
https://x.com/wayve_ai/status/2095406780364787937
BBC Boss Matt Brittin Outlines Concept For “Public Service Algorithm”
https://deadline.com/2026/08/matt-brittin-bbc-iplayer-overhaul-public-servic-algorithm-1237049205/
BBC’s director general Matt Brittin has said BBC iPlayer could benefit from a “public service algorithm” that allows it to function more like Youtube and opens up to creators
https://x.com/pressgazette/status/2093235516485321142
A ottobre porto a Torino un piccolo gruppo dell’ecosistema robotico italiano. Durante Wave byVento (ex Italian Tech Week) ospito un side event @22Astronauts_: Costruire il futuro dell’industria italiana con la robotica Torino, 8 ottobre L’idea è semplice: riunire founder,
https://x.com/IlirAliu_/status/2095210396613304634
A robot in Tübingen is learning to pick tomatoes by watching humans. A few days ago, I spoke with Martin Kiefel, co-founder of polybot. Ahead of @Cyber_Valley 10th anniversary on September 9, I’m spending time with some of the people building the next generation of AI and
https://x.com/IlirAliu_/status/2094334877982461979
AI systems are impressive until something happens that was not in the training data. Next week @Cyber_Valley turns 10, and I am helping them to showcase people and companies from the ecosystem, that are working on todays problems. I recently sat down with Johannes Haux,
https://x.com/IlirAliu_/status/2095572785233866893
Germany produces around 86 million tonnes of construction and demolition waste every year. A surprising amount of the quality control still relies on samples, visual inspection and someone making a judgement call about what is actually sitting in front of them. @optocycle is
https://x.com/IlirAliu_/status/2095134899749753292
X Square Robot has introduced TwinDEX: a co-designed pair of three-finger, 9-DoF dexterous hands, one wearable for data collection, one matched robot end effector for deployment. The wearable and the robot hand share the same kinematics, geometry, materials, look, and sensors,
https://x.com/TheHumanoidHub/status/2095178516639969778
Leave a Reply