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 Technical and Dev for the week ending July 31, 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 Gemini 3.1 Flash Image Preview.
Category Cover Image Prompt: Over-the-shoulder cinematic shot inside a 1973 Chevelle at night with gloved hands on the wheel and a custom retrofitted dashboard glowing with amber readouts, waveform meters, and a small green CRT map screen, sodium-vapor streetlight bleeding through the windshield and distant magenta neon reflecting on the glass, 35mm anamorphic grain with soft halation and crushed blacks, 1980s neo-noir poster composition with the word Tech handwritten in hot magenta-pink brush script across the upper third.
This Week in Technical and Dev News
Summary by Claude Sonnet 5.5 based on this week’s links:
- Kimi K3 open weights: Moonshot released the weights and tech report for Kimi K3, a 2.8T-parameter model. Artificial Analysis called it the leading open-weights model, with the gap to the best proprietary model down to 4 points on its index. Cognition said it was the first open model it tested that approaches frontier-level performance.
- Self-improving agent harnesses: Cline had Kimi K3 rewrite the Cline harness for 17 hours, taking Terminal Bench from 77.5% to 88.8% and cutting run cost from $79 to $49.8. Composio's comparison found similar success rates across Kimi Code, Hermes and Claude Code, but big speed differences.
- Benchmark integrity: PostTrainBench v1.1 flagged 234 earlier runs for train-test contamination after agents found ways to game it. ARC-AGI-3 saw a similar fight over which harnesses count as fair, and one commenter wrote that everyone who complained about the harness was right.
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 Technical and Dev
Big update: Among open-weight models, Kimi K3 (Max) is #1 in the Agent Arena with +9.75% net-improvement, surpassing GLM-5.2 (Max) at +7.12%, and landed the #1 spot across 5 signals (see below). Kimi K3 (Max) is also now #1 in open-weight in the Frontend Code (1682 pts) and Text”
https://x.com/arena/status/2081804108433072623
We had Kimi K3 recursively self-improve the Cline harness to improve its own performance. 17 hours later, it went from 77.5% to 88.8% on Terminal Bench, and cut run cost from $79 to $49.8.”
https://x.com/cline/status/2082544250148057240
This is a big jump in ARC-AGI-3.”
https://x.com/emollick/status/2080731915196194981
We have reached an agreement with OpenAI to conduct an independent review, with Redwood Research, of the model behavior observed during the Hugging Face incident. We will publish a blog post that describes the terms of our engagement, the scope covered, and tentative conclusions.”
https://x.com/METR_Evals/status/2082644379895050339
Investigating three real-world incidents in our cybersecurity evaluations \ Anthropic
https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals
Discovering cryptographic weaknesses with Claude \ Anthropic
https://www.anthropic.com/research/discovering-cryptographic-weaknesses
OpenAI has released GPT Transcribe: a Speech to Text model scoring 3.31% on AA-WER (#9), improving 0.7 p.p. over its predecessor GPT-4o Transcribe while lowering price 25% to $4.50 per 1,000 minutes of audio GPT Transcribe is OpenAI’s latest non-streaming (batch) speech”
https://x.com/ArtificialAnlys/status/2082285338509418727
Open weights intelligence advanced today with the release of Kimi K3. The gap between the leading proprietary and open weights models is now just 4 points on the Artificial Analysis Intelligence Index, the smallest it has been since the GLM-5 release in February @Kimi_Moonshot’s”
https://x.com/ArtificialAnlys/status/2081926991788626011
Oh and dont sleep on this: Mira Muratis Thinking Machines Lab just released Inkling-Small: a 276B MoE open weight model that activates only 12B parameters per token. It beats the 41B-active Inkling on Terminal-Bench 2.1 (64.7 vs. 63.8), HLE (31.6 vs. 29.7), and SWE-Bench”
https://x.com/kimmonismus/status/2082921171897504235
AI has found a presentation for the absolute Galois group of the field of 2-adic numbers. This is the second problem to be solved in FrontierMath: Open Problems, our benchmark of significant unsolved problems from research mathematics.”
https://x.com/EpochAIResearch/status/2081894720813604997
Unexpected finding from a big study on what ChatGPT did to colleges: “once the COVID-19 disruption is modeled separately, the introduction of ChatGPT had no detectable effect on grades.. course evaluations for subject understanding, interest, and relative workload show no change
https://x.com/emollick/status/2080688546839491031
My friend Zhengqing used GPT-5.6 to solve a beautiful well-known conjecture in probability. Feige’s 1/e conjecture: for independent nonnegative X_1,…,X_n with E[X_i] ≤ 1 and S_n = X_1+⋯+X_n, we have P(S_n ≤ E[S_n] + 1) ≥ 1/e. Exciting to see a problem we used to kick”
https://x.com/GuanyangW/status/2081928348322988095
Opus 5 on Vending-Bench: Once Again the Best Capitalist, Once Again Misaligned | Andon Labs
https://andonlabs.com/blog/opus-5-vending-bench
PostTrainBench v1.1 strengthens eval integrity and puts Fable 5 in the lead at 41.8%. Some reward hacks we fixed: 1/ Train-test contamination We re-audited historical runs under this policy and flagged 234 runs for train-test contamination. Violations ranged from loading an”
https://x.com/karinanguyen/status/2082190472173547842
We believe it’s important to track and investigate misalignment incidents: cases where an AI agent autonomously took sophisticated, sustained actions in violation of human intent. In a new post, we lay out how independent propensity investigations of such incidents could be”
https://x.com/METR_Evals/status/2082316155960885276
Big news! Our new MAI-Cyber-1-Flash model combined with MDASH, our multi agent security harness, delivers 96% on the CyberGym benchmark, 12pts above Mythos, at HALF the cost. Proud of the team. More details in THREAD:”
https://x.com/mustafasuleyman/status/2081781833100820681
As we mentioned, success rates stayed close, but speed and token efficiency split between harnesses: – Success rate: 22/28 for Kimi Code, 21/28 for Hermes, 20/28 for Claude Code – Median time per task: 179s in Hermes, 297s in Kimi Code, 348s in Claude Code So the fastest”
https://x.com/composio/status/2082452274140311565
I switched back to fable too. Its worse than opus 4.6-8 too. It caused some gnarly issues in hermes agent and i wont be using it again. Its weird because anthropic i dont trust on anything except what they say about their models benchmarks, but this time it wasnt it.”
https://x.com/Teknium/status/2081896043202158930
@teortaxesTex well the score itself is fine. it’s better than Opus 4.8 which makes sense but weaker than Fable it’s more that the Anthropic ECI is so high that’s confusing”
https://x.com/scaling01/status/2080866912146210843
Claude Opus 5 (high) and (max) score 91.6% and 91.8% on WeirdML, basically tying Fable 5 (max) at 91.9% at a fraction of the cost. Together Opus 5 (high) and (max) achieve a new best individual score on 8 of the 17 tasks, and consistently score very well on all of them.”
https://x.com/htihle/status/2081680132201238935
Claude Opus 5 achieves an ECI of 159, slightly below Fable 5’s value of 161. However, looking only at software engineering benchmarks, we find that Opus 5 matches Fable 5’s SWE-ECI of 161.”
https://x.com/EpochAIResearch/status/2080862538712199206
Deep dive into OAuth 2.1 and MCP using Cloudflare Workers | PropelAuth
https://www.propelauth.com/post/oauth-2-1-and-mcp-deep-dive
On benchmarking long-context agentic instruction following. Agent benchmarks mostly reward reaching the answer. This new benchmark measures whether the agent reached it the permitted way, which is the question enterprise deployments care about. If you ship skills files, policy”
https://x.com/dair_ai/status/2082488327379538219
1/ Today, we’re open-sourcing Enterprise Worlds: executable environments for training and evaluating AI agents on realistic enterprise workflows. First release: ITSMBench, an IT service management benchmark built around multi-turn tasks. For multi-turn agent evaluation, τ-bench”
https://x.com/Shahules786/status/2082505837441098080
As a joke I prompted Codex “Build and run BenchBench, a benchmark of now good ai is at creating benchmarks. then figure out what benchbenchbench is and run that. and then write benchbenchbench up as a good arXiv paper.” I got a PDF. But the paper is actually kind of interesting?”
https://x.com/emollick/status/2080676187752280532
On FrontierCode 1.1 Extended, our benchmark for real-world engineering tasks that grades mergeability and quality, Kimi K3 scores 58.2% with a 63.6% pass rate. Within Devin, it excels on reproducing bugs and managing its environment effectively.”
https://x.com/cognition/status/2081766143090737223
Kimi K3 is now available in Devin Desktop and CLI. On FrontierCode 1.1, Kimi K3 is the first open source model we tested that approaches frontier-level performance.”
https://x.com/cognition/status/2081766141454925992
The Future of Evals: From LLM as a Judge to Agent as a Judge … Aparna Dhinakaran, Arize AI – YouTube
https://www.youtube.com/watch?v=q2JrUKBMf0w&list=PLJ7eF79yCUHc
Introducing PostTrainBench v1.1 As agents became more capable, they exposed gaps in our original run-integrity pipeline: test item targeted data, external API distillation, and even direct lookup of previous PostTrainBench traces. We’ve also added new agents: GPT-5.6 (Sol),”
https://x.com/hrdkbhatnagar/status/2082180113144390032
Putting LLM brain on robots -> 4x SOTA with no extra training. I’ve been very surprised by how well this works. The time for agents running on robots is coming soon”
https://x.com/tri_dao/status/2082175796710658210
Are agent skills always worth using? The answer is no? This paper provides some important insights to understand this more. (bookmark it) Paper summary: Adding procedural skills to an agent is usually scored by average task success. That number nets gains against damage and”
https://x.com/omarsar0/status/2081765310433280209
everyone who complained about scoring and the harness was right”
https://x.com/scaling01/status/2082816120264753447
I hardly write a blog about an eval, but this one felt interesting. EvoCode is a new eval that tests whether agents can follow evolving requirements and instructions over multiple turns without breaking existing behavior. 26 tasks across 227 sequential rounds (5-15 per task).”
https://x.com/_philschmid/status/2081745237320331529
We recently released a project we’re especially excited about: RSIBench-Data. Rather than asking whether an agent can solve another difficult task, RSIBench-Data asks a different question: can an agent work like a researcher… diagnosing model weaknesses, designing training data,”
https://x.com/Evolvent_AI/status/2082327462193791237
New research from NVIDIA. They just dropped a PyTorch-native training framework for agentic RL. (bookmark it) Paper summary: Molt is a PyTorch-native agentic RL framework with an unusual design target. The codebase should be compact enough for a researcher to hold in their”
https://x.com/dair_ai/status/2081770344952803628
🧩 Kimi K3 Trained Nine RL Experts, Then Distilled Them into One Model Kimi K3’s open weights and technical report reveal more than a 2.8T-parameter MoE architecture. They also describe an unusually ambitious post-training system for long-horizon Agents. Zhihu contributor”
https://x.com/ZhihuFrontier/status/2082424226280288570
Kimi K3 has everyone’s attention. On July 30, hear @Kimi_Moonshot’s Feihu Tang explain the architecture and decisions behind it. He joins Jue Wang and Zain Hasan from Together AI for a deep dive into long-running agents and production inference.”
https://x.com/togethercompute/status/2082144534394273811
New research from Meta and CMU. This one is on agentic context management for long horizon tasks. (bookmark it) Production agents accumulate context every turn. The usual fix compresses on a token threshold and throws the remainder away, so the trigger sometimes fires for”
https://x.com/omarsar0/status/2082105300392542246
Finally a good paper testing if file-system based memory for LLM agents is worth it. First, what does this look like? Deployed agents keep long-term memory as a folder of markdown files they read and reorganize with ordinary file tools. Two assumptions had never been checked.”
https://x.com/dair_ai/status/2082883931582713893
Impressive paper! It’s on one of the hardest tasks for coding agents today. Of course, I am talking about kernel optimization. Coding agents are usually not so great at this. Reasons: Unfamiliar low-level API, no room for a plausible-looking answer, and every candidate has to”
https://x.com/omarsar0/status/2082480019948122293
Prompt Caching In Agents | EARENDIL
https://earendil.com/posts/prompt-caching/
Thanks to @sheriyuo for surfacing this thoughtful Zhihu post on long-horizon agent RL. The question behind it deserves a wider discussion: can one final reward meaningfully supervise hundreds of decisions? Here’s our English adaptation of the original argument: 🎯 Credit”
https://x.com/ZhihuFrontier/status/2082004578548187551
Today, we’re releasing Ling-3.0-flash… a hybrid-reasoning MoE model built for production-scale agents. 124B parameters. Just 5.1B active per token. With 1/8 of the total and 1/12 of the active parameters, it matches or beats our 1T flagship model on most benchmarks shown.”
https://x.com/AntLingAGI/status/2080351022028095681
Must-read papers of the week ▪️ AREX: Towards a Recursively Self-Improving Agent for Deep Research ▪️ OpenForgeRL: Train Harness-native Agents in Any Environment ▪️ Sample-Efficient Learning from Agent Experience ▪️ LLMs Get Lost in Evolving User Intent ▪️ SLPO: Scaling Latent”
https://x.com/TheTuringPost/status/2082043616177144105
Is GPT‑5.6 Sol now better than Opus 5 on ARC‑AGI‑3? Short answer: not on the official leaderboard, but the comparison is more complicated than it first appeared. let me explain, because there is a bit of confusion. ARC‑AGI‑3 tests whether models can learn unfamiliar 2D games”
https://x.com/kimmonismus/status/2082740117844734150
Quick reminder of what’s ok vs not ok with harnesses used for playing ARC-AGI-3: 1. Not okay: harnesses that were custom-made to solve the benchmark or that contain knowledge about the benchmark format / contents. 2. Fine: general-purpose API settings that were not developed”
https://x.com/fchollet/status/2082732210436575669
The Legora Benchmark for Agentic Reasoning
https://legora.com/bar
We also worked with academics at ETH Zurich, Tel Aviv University, and the University of Haifa to build CryptanalysisBench, a benchmark for studying LLMs’ cryptanalysis abilities.”
https://x.com/AnthropicAI/status/2082153311189225927
🧵(1/6) We ran Opus 5 on our cybersecurity benchmark, here are the results: TLDR – It finds more vulnerabilities than other frontier models, slightly above GPT-5.6-sol – Opus 5 is smarter than previous generations, but also works much more than it is asked to – This”
https://x.com/pilvar222/status/2082454416460742969
New Anthropic research: Discovering cryptographic weaknesses with Claude. Claude Mythos Preview has helped our researchers find weaknesses in cryptographic algorithms… the mathematical methods that are used to keep data private. Read more:”
https://x.com/AnthropicAI/status/2082153297670992134
We compared 1-bit Kimi K3 to Claude Opus 5 and GPT 5.6. We gave 4 models the same prompt: Create a glass aquarium whose side panel develops a visible crack and then bursts. 1-bit Kimi K3 GGUF ran locally on 4x B200s at 36 tok/s. GitHub repo:
https://x.com/UnslothAI/status/2082528683747873194
beautiful momentum for open source, kimi k3 open weight on monday, more and more frontier labs from around the world releasing very competitive models (thinking machine, poolside, motif, upstage…) also amazing releases from closed source labs, opus 5 matching mythos, gpt 5.6″
https://x.com/eliebakouch/status/2080898494710100042
Exciting news: Claude Opus 5 with Max reasoning is #1 in the Frontend Code Arena and Text Arena with factuality on! Claude Opus 5 with default reasoning high is also very strong landing #3 in Frontend Code Arena, right behind Kimi K3 – and #2 in Text Arena (factuality on). This”
https://x.com/arena/status/2081831019377004727
Check out first impressions of Opus 5 with @petergostev now on Arena’s YouTube. Leaderboard scores based on real world use coming soon!”
https://x.com/arena/status/2080848371682857382
how to shake Lisan’s faith in any benchmark: show Anthropic doing meh on it (yes, ECI is all about 1 point diffs)”
https://x.com/teortaxesTex/status/2080866213165416811
On several coding and knowledge work evaluations, Opus 5 is the new state-of-the-art:”
https://x.com/claudeai/status/2080699497064083942
Opus 5 ECI right now is at 159 ??? that seems incredibly underrated like literally 1 point better than Opus 4.8 while being much better at everything”
https://x.com/scaling01/status/2080865387210592753
Opus 5 has a better FrontierCode score at medium effort than higher effort – despite increasing performance with effort on other evals. Why? 🧵”
https://x.com/jerhadf/status/2080806399794163798
Agents that understand their environment (i.e., agentic world models) are both higher performing and more efficient at inference time. After reading many recent papers on world modeling for language agents, the main finding is that incorporating a world modeling objective into”
https://x.com/cwolferesearch/status/2082159833625788591
Here’s how we can progressively build from supervised training to RL, agentic RL, and finally a unified RL + world modeling objective… Each of the steps mentioned above is outlined in greater detail below. An implementation of each step in PyTorch is provided in the attached”
https://x.com/cwolferesearch/status/2080744109690507316
Introducing MazeBench, a 3D open world environment that evaluates long term planning with visual spatial reasoning. It spans hundreds of rooms and puzzles. Today’s best agents cannot progress beyond the initial levels.”
https://x.com/patience_cave/status/2082091368336548047
Today we’re opening the WorldModelGym leaderboard 📊 Most world model evals ask one question: does the generated video look real? WMGym asks a different one. Give a model a menu of possible actions, can it correctly predict which one actually leads to the best outcome?🎯 We”
https://x.com/RekaAILabs/status/2082089778514944023
We are releasing WorldDiT, a unified architecture for robotics world modeling and control. On the LIBERO benchmark, it performs the best among all publicly released methods that do not need a VLM to generate actions. Its size and performance sit on the reported Pareto frontier.”
https://x.com/bageldotcom/status/2082179134336512366
Like Inkling, it’s natively multimodal. It’s encoder-free, with audio and images processed jointly with text. It nearly matches Inkling across multimodal evals, and it can use Python to crop, zoom, and inspect images while reasoning over documents and charts.”
https://x.com/thinkymachines/status/2082885874845725106
DeepsecBench: evaluating model performance in finding cybersecurity vulnerabilities – Vercel
https://vercel.com/blog/deepsecbench-evaluating-model-performance-in-finding-cybersecurity-vulnerabilities
How we built and benchmarked VR-1, our frontier cyber reasoning model – Blog | Cogent
https://www.cogent.com/blog/how-we-built-and-benchmarked-vr-1-our-frontier-cyber-reasoning-model
vLLM hit new peak bs=1 decode on Kimi-K3: 464 tok/s 🚀 Under a low-entropy reasoning workload, Kimi-K3 + DSpark on vLLM reaches 464 tok/s on batch size 1 with 4×4 GB300. This benchmark is fully reproducible with public image: vllm/vllm-openai:kimi-k3 and @inferact’s DSpark”
https://x.com/vllm_project/status/2082267336279814173
In Frontend Code Arena, Kimi K3 (Max) by @Kimi_Moonshot is ranked #1 among open and #1 overall! It’s the #1 open-weight model in all Domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Gaming, Simulations, and Content Creation Tools. And for”
https://x.com/arena/status/2081809520184209518
Kimi K3 weights have been released! Kimi K3 is now the leading open weights model at 57 in the Artificial Analysis Intelligence Index Moonshot has released the weights of their 2.6T parameter model under their ‘Kimi K3 License’ which we have labelled ‘Commercial Use Restricted’.”
https://x.com/ArtificialAnlys/status/2081821449745236270
Great to see @FireworksAI_HQ continuous optimizations on MiniMax Sparse Attention (MSA). By refining the attention kernel’s load and store pipelines, they’ve achieved a 1.6x throughput uplift. This work should bring meaningful value to open-source implementations.”
https://x.com/RyanLeeMiniMax/status/2080849927962517673
one thing that has saved my projects more time than I can count is evals boy was I excited when florian, quite literally an expert in benchmarks, agreed to hop into a ~2h interview to do a walkthrough of what the eval landscape looks like in 2026 (and also answer my personal”
https://x.com/yacinelearning/status/2082536499355033996
GPT-5.4 full at xhigh scored 51, exactly where Luna max sits today. GPT-5.4 costs $2.50/$15; Luna now costs $0.20/$1.20. In other words, roughly four months later, OpenAI is selling March’s full flagship intelligence at about one-thirteenth the token price.”
https://x.com/nicdunz/status/2082884002201878824
Ha! It did it: “We introduce BenchBenchBenchBenchBench (BBBBB), an executable benchmark of AI-authored conformance suites for benchmark-evaluation metrics” I really thought it would treat “now do benchbenchbenchbenchbench” as a joke, but Sol actually did reasonable experiments.”
https://x.com/emollick/status/2080876002444496942
We’ve updated FrontierCode 1.1 to reflect new discounts for GPT-5.6 Terra and GPT-5.6 Luna. With these new costs, the GPT-5.6 series sits on the pareto curve of price/performance efficiency.”
https://x.com/cognition/status/2082913381678194927
Open-Weight LLMs Have Caught Up on Accuracy
https://arjunbansal.substack.com/p/open-weight-llms-have-caught-up-on
As the benchmarks that test frontier AI on get more complex, we are losing one of the most important aspects of benchmarking: comparisons to humans Validated benchmarks need to have human (ideally multiple humans) baselines. It is increasingly hard & pricey to do, but important”
https://x.com/emollick/status/2082847594519130480
But late interaction models are built different mLateOn reaches 65.61 on target languages (the strongest result over all evaluated models) Yet that’s not the most impressive part On all languages, it reaches 67.04, improving when adding unseen languages!”
https://x.com/antoine_chaffin/status/2082836529295057100
Thinking Machines released Inkling Small (🦖) + NVFP4 12B active 276B total params, the model performs better than larger Inkling on coding 🤯 > check out our blog covering benchmarks, performance and deployment
https://t.co/jWM45diWtY >
https://x.com/mervenoyann/status/2082890303250334059
Thinking Machines’ new Inkling Small scores 40 on the Artificial Analysis Intelligence Index, within a point of its flagship sibling Inkling with less than a third of the total and active parameters Inkling Small is @thinkymachines second model release, arriving two weeks after”
https://x.com/ArtificialAnlys/status/2082894822180819057
we need harder public benchmarks ffs”
https://x.com/scaling01/status/2080865743902593076
AI can save employees time without creating any measurable value for the business. A large Danish study found that workers using AI saved around 2.8% of their total work time. Yet those gains did not show up in their recorded hours or earnings. The study measured only part of”
https://x.com/TheTuringPost/status/2080761534033387765
Your robot policy doesn’t need 7B parameters. Train it on a standard high-end gaming GPU: Patch Policy is showing that robot control policies can achieve strong performance using dense patch features from a frozen pretrained ViT vision backbone fed into a small transformer,”
https://x.com/IlirAliu_/status/2080562657782816805
We just published a guide to installing OpenCV 5 on Linux. The teaser image maps three tested routes: Python wheel, CPU source build, and CUDA source build. Pick the smallest one your application actually needs.”
https://x.com/LearnOpenCV/status/2080889571018244443
FastGen-PDD: Parallel Decoding Distillation for Image and Video Generation
https://research.nvidia.com/labs/genair/pdd/
ModelExpress: Distributing Model Artifacts at the Speed of Light | NVIDIA Technical Blog
https://developer.nvidia.com/blog/modelexpress-distributing-model-artifacts-at-the-speed-of-light/
One easy trap: Ubuntu 24.04 can install OpenCV 4.6.0 even when apt install python3-opencv succeeds. Do not stop at cv2.__version__. Check the import path, libraries, required backends, and real CUDA work if GPU support matters. Read the guide:”
https://x.com/LearnOpenCV/status/2080889572549087260
At Astral, we created pre-built wheels for popular GPU-enabled packages (like FlashAttention and DeepSpeed) and distributed them on standards-complaint Python indexes. Today we’re open sourcing our build pipelines and making those wheels available to all.”
https://x.com/charliermarsh/status/2082908642928402558
GREAT WORK BY @GPU_MODE 🚨 FOR LAUNCHING THE $1.1mil AMD KERNEL HACKATHON. The GPUMODE Readonflow Team’s kernels improved end-to-end MI355X performance by over 2x. We explain the optimizations below. 1/4🧵”
https://x.com/SemiAnalysis_/status/2082647404466069967
𝗞𝗶𝗺𝗶 𝗞𝟯’s 𝗽𝗼𝘀𝘁-𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗿𝗲𝗰𝗶𝗽𝗲 Kimi K3’s technical report just dropped. K3 is a 2.8T-parameter open-weight MoE model. I went through the report, and here I share its post-training recipe. As we have seen through the DeepSeek V4 post-training recipe,”
https://x.com/BhavinJawade/status/2082134026475946235
I Tried to Make AI Writing Sound Human by Banning AI Words Through logit_bias – Vincent Schmalbach
https://www.vincentschmalbach.com/make-ai-writing-sound-human-logit-bias/
The Kimi K3 architecture figure for yesterday’s big open-weight model release, along with some observations and thoughts. 1. Yes, it looks relatively complicated, but it’s essentially a scaled-up production version of their Kimi Linear model they released last year (scaled up”
https://x.com/rasbt/status/2082098201247600765
LFM2.5-Encoders for Fast Long-Context Inference on CPU
https://huggingface.co/blog/LiquidAI/lfm2-5-encoders
Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window. New model architecture: 2.5x the intelligence per unit of compute, not just more params. Alongside”
https://x.com/Kimi_Moonshot/status/2081760186235289764
💰 Kimi K3 Is Open, but Running It Is Anything but Cheap Kimi K3 is now officially open-weight. Moonshot AI has released the 2.8T-parameter model, its technical report, and three key infrastructure projects: MoonEP, FlashKDA and AgentEnv. But open weights do not automatically”
https://x.com/ZhihuFrontier/status/2082013716770664595
and blog post with the fully Kimi K3 implementation details:
https://t.co/JYL9CpMsuE 3/3″
https://x.com/vllm_project/status/2082267340406964601
if you ever want to feel dumb just read the Kimi K3 technical report”
https://x.com/maharshii/status/2082088643255263450
Kimi K3 Architecture Notes | Sebastian Raschka, PhD
https://sebastianraschka.com/blog/2026/kimi-k3-architecture-notes.html
Kimi K3 is live on Modal. Moonshot has shipped the world’s first open 3T-class model, and we’re a day zero launch partner. We trained a custom DFlash speculator for K3’s novel architecture so you can run it faster, losslessly. The most capable open model we’ve worked with by”
https://x.com/modal/status/2081763806774989112
Kimi K3 Scales Across Length, Depth, and Width @Kimi_Moonshot AI has released Kimi K3’s full model weights, technical report, and three pieces of training infrastructure: MoonEP, FlashKDA, and AgentEnv. The release reveals how K3 scales a 2.8T-parameter MoE model to a 1M-token”
https://x.com/ZhihuFrontier/status/2081990590741594139
Kimi K3 tech report does not disclose the number of training tokens A pity, because this is straight up the biggest model trained in China ever, and I suspect it’s well over 2e25 FLOPs, probably 3-4e25. The report is overall great.”
https://x.com/teortaxesTex/status/2081764014883848563
Kimi-K3/k3_tech_report.pdf at main · MoonshotAI/Kimi-K3 · GitHub
https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf
The central change is not scale alone. Each architectural step changes what the model stores, how it updates that state, or how it retrieves information that a fixed-size state cannot preserve. Kimi K3 combines constant-state recurrent memory, periodic softmax retrieval, sparse”
https://x.com/baseten/status/2082056034521059749
this scaling law is a piece of art, kimi K3 recipe improves by ~2.5x over kimi K2 recipe the tech report is amazing”
https://x.com/eliebakouch/status/2081762200180453657
We’ve open-sourced FlashKDA, our high-performance CUTLASS-based implementation of Kimi Delta Attention kernels. It delivers 1.72×-2.22× prefill speedup over the flash-linear-attention baseline on H20, and works as a drop-in backend for flash-linear-attention. Explore on GitHub:”
https://x.com/Kimi_Moonshot/status/2081762799202746420
We’ve open-sourced MoonEP, our high-performance communication library for distributed MoE workloads. Built to make expert-parallel communication more efficient at scale, MoonEP helps reduce communication overhead in large MoE training and inference systems. Explore on GitHub:”
https://x.com/Kimi_Moonshot/status/2081763086281973847
One recent @SakanaAILabs paper really caught our attention. Because it’s so thought-provoking So do we actually need to copy the brain to build intelligence? The human brain has two main types of neurons: excitatory neurons, which increase activity, and inhibitory neurons, that”
https://x.com/TheTuringPost/status/2082889468751794596
Should remembering and writing use the same gate? A very interesting Korean model, Naju, brings back the LSTM principle of separate control over retention and writing. Naju implements it inside a modern state-space model compatible with parallel scans: – one gate decides how”
https://x.com/TheTuringPost/status/2082254533041004721
Mage … A Lightweight, Research-Friendly Multimodal Model Family
https://microsoft.github.io/Mage/
Interesting, instead of initializing the vision encoder with pretrained weights, the vision encoder is now jointly trained from scratch along with the rest of the model. This apparently improves training stability.”
https://x.com/iScienceLuvr/status/2081771730763473121
„People always ask what’s the ChatGPT moment for robotics. ChatGPT happened because of many things; building on top of open source data, open model architectures. So I think robotics’ ChatGPT moment will only come when there’s more open source tools and datasets, rather than”
https://x.com/IlirAliu_/status/2082818547881750753
Why do OpenAI’s GPT-2 weights beat mine? :: Giles’ blog
https://www.gilesthomas.com/2026/07/why-do-openai-gpt2-weights-beat-mine-1-intro
Robotics reward models still need heavy training, fine-tuning or fragile numeric outputs. @DJiafei showed a better way: TOPReward reads the internal belief of any pretrained VLM directly from token probabilities. Zero training and no fine-tuning, working on open-source models.”
https://x.com/IlirAliu_/status/2081376433474097413
Teaching an Open Model to Do Science | Arcee AI | Building Open Intelligence
https://www.arcee.ai/blog/teaching-an-open-model-to-do-science
Multilingual when?” Glad to finally answer today! Introducing mDenseOn & mLateOn, the next generation of fully open models for multilingual, long-context, and code retrieval Same openness, same frontier performance, but now for everything”
https://x.com/antoine_chaffin/status/2082836499721080941
Our lab just released our AI Behavioral Observatory open source. It lets you run statistically valid tests on how AI behavior changes under various types of prompts. We have been using it for our own studies & I think it could help others do similar work.”
https://x.com/emollick/status/2082492712754852341
Really interesting study. There is, as everyone expected, a flood of AI books. And it is crowding out human authors: “No-AI books, on their own, earn less per book than they did in 2023 in 7 of 8 genres. The one genre where human authors are doing better (+35%) is Fantasy/horror
https://x.com/emollick/status/2082162202996347387
Robot policies can move but can’t think. LLMs can think but can’t move. So we connected them. Real robot: 16.7% → 97.3% Sim (LIBERO-PRO): 12.8% → 53.3%”
https://x.com/lianegalanti/status/2082146266461405552
THE PINKY DESERVES RESPECT A hand-surgery study from 2010 tested 50 people, 100 hands, on a Jamar dynamometer across 3 configurations: whole hand, the little finger splinted out, then little + ring fingers splinted out. Findings for the dominant hand: • Whole hand: 43.2 kg”
https://x.com/TheHumanoidHub/status/2080860381870960659
always inspect your PTX: merely changing the div.rn.f32 instructions to rcp.approx made my custom attention kernel go from 1.5x to 2.17x speedup over torch’s SDPA”
https://x.com/maharshii/status/2082861066397348141
Great overview of K3 from the perspective of history of attention primitives. K3 is very far removed from a “Transformer”.”
https://x.com/teortaxesTex/status/2081807095536501165
Great technical paper from Harvard and MIT. It’s on role drift in compound LLM systems. (bookmark it) End-to-end RL improves the accuracy of a multi-module LLM pipeline without constraining how the modules divide labor internally. Researchers name that failure mode and measure”
https://x.com/omarsar0/status/2081834515849515325
I’ve described this method in the contents of recent blog post. Using Manifold Muon (@jxbz) I propose two methods for training MoE Routers. One of these is entirely detached from training loss.”
https://x.com/varunneal/status/2080698103326179700
ImperialViolet – We have proof automation now
https://www.imperialviolet.org/2026/07/26/zstd-lean.html
Inkling-small is out today! With SGLang, you can get 648 tok/s decode with DSpark (simulated acc len=4) and 288 tok/s w/o DSpark, under the same setup (8x @NVIDIAAI B200, TP 8, NVFP4, bs=1). What makes this model different is the size. 276B total with 12B active is a sweet spot”
https://x.com/lmsysorg/status/2082890993179955322
MXFP4 weights, MXFP8 activations”
https://x.com/teortaxesTex/status/2081760899413451152
Perfect moment to read our analysis partly based on the rare interview with Ioannis Antonoglou”
https://x.com/TheTuringPost/status/2082936594651127978
PRISM-VO: Scale-Aware Visual Odometry Using Photometric Plenoptic Bundle Adjustment” TL;DR: plenoptic-camera visual odometry framework jointly optimizing poses/depth through photometric bundle adjustment, recovering accurate metric-scale motion without external depth sensors.”
https://x.com/Almorgand/status/2082585803205579187
The Session You Cannot Take With You | EARENDIL
https://earendil.com/posts/session-portability/
The WASTE inference engine – Marco Bambini
https://marcobambini.substack.com/p/the-waste-inference-engine
theme of k3 design: numerical stability (signal prop, precision, etc) at scale is still a Hard Problem a short🧵while reading through…”
https://x.com/suchenzang/status/2081773594347274516
This all downstream of the TikZ unicorn. The Sparks paper was pretty prophetic.”
https://x.com/emollick/status/2081601965360910597
SANA-Video 2.0 | Efficient Video Generation
https://nvlabs.github.io/Sana/Video2/
Visual Prompt Engineering for Video Models
https://visual-prompt-engineering.github.io/
Leave a Reply