🚨 Cursor's MoE megakernel, Qwen3.8-Max at 2.4T, and a charting library that ships 10M points in 258 KiB

Presented by Mend.io  Securing AI Agents, MCP Servers & LLM Apps — agent behavior is not defined by code alone. Get the framework →

Every layer below used to be lossy at the boundary. HTML parsing drops layout. Re-serializing a .docx drops formatting. Per-row draw calls force you to decimate before you plot. A single aggregate accuracy drops the failure mode. All eight releases move the same way: keep the canonical representation, emit a diff against it, and surface the loss where one is unavoidable.

2.37×

MoK MXFP8 forward speedup

258 KiB

XY export, 10M points

2.4T

Qwen3.8-Max total parameters

01 · Cursor · MoE training

Cursor Open-Sources Mixture-of-Kittens (MoK): A Deterministic MoE Training Megakernel for GB300 NVL72 Racks

The MoE layer used to be a compute problem. On a 72-GPU NVLink domain it became a scheduling problem, and the MoE layer can consume more than half of end-to-end training time. Cursor fused every communication and computation step into one deterministic megakernel and took the CPU out of the loop.

Pull-based dispatch plus push-based combine cut dispatch signalling from 103 µs to 18 µs, roughly 5.8x. Against the fastest public baseline it runs up to 2.37x for MXFP8 forward, 1.78x MXFP8 backward, 1.92x BF16 forward, 1.58x BF16 backward. End to end on 512 GPUs, tokens per second per GPU moved from 760.9 to 1,070.2. Apache-2.0, but the floor is Blackwell SM100/SM103, CUDA 13.0+ and PyTorch 2.10+.

Read on Marktechpost →  · GitHub  · Technical details

02 · Alibaba Qwen · Frontier model

Alibaba Qwen Releases Qwen3.8-Max: A 2.4 Trillion Parameter MoE Model and the Most Capable One in the Qwen Family to Date

A flagship release usually arrives with one number you are meant to remember. This one arrives with a number that is missing. Alibaba has not disclosed the activated-parameter count, so serving cost for the 2.4T checkpoint cannot be modeled yet. Open weights for Qwen3.8-Max and Qwen3.8-27B are promised next week.

The model page lists a 1M-token context window, with maximum input at 991K tokens and 983K when thinking is enabled. Pricing is $2.00 per 1M input tokens and $6.00 per 1M output, with implicit cache reads at $0.25 — cached input is eight times cheaper than fresh, so prefix stability drives cost more than prompt length. It scores 86.6 on Terminal-Bench 2.1 and 92.6 on GPQA Diamond, but reports 67.7 on SWE-bench Pro and 73.5 on FrontierSWE against Fable 5’s 80.0 and 88.8. The gains are multimodal and agentic, not reasoning.

Read on Marktechpost →  · Announcement  · Model page  · Qwen Studio

03 · Reflex · Python charting

Reflex Open Sources XY: A Rust-Backed Super-Fast Python Charting Library That Keeps 100 Million Point Charts Interactive

Every Python charting stack makes one drawable object per row, so past a few hundred thousand points you start sampling before you plot. XY draws a screen-bounded representation instead and leaves the exact f64 columns in Python, so hover, selection and zoom drilldown still return original rows.

Render time holds at 0.071 s for 10,000 points and 0.081 s at 100 million. At 10M the benchmark reports 0.083 s against 2.804 s for Matplotlib WebAgg and 3.367 s for Plotly scattergl; Plotly does not complete at 50M. A 10-million-point interactive scatter exports to 258 KiB of HTML against a stated 259 MiB for the Plotly equivalent, and the payload stays near 258 KiB from 1M through 100M rows. Apache-2.0 — and version 0.0.1 early alpha, which argues against critical paths.

Read on Marktechpost →  · GitHub  · Benchmarks  · Blog

Sponsored · CopilotKit

Your agent already works. Channels SDK gives it somewhere to work with people.

CopilotKit has released the Channels SDK, an open source library that puts an existing agent inside Slack and Microsoft Teams without a rewrite per platform. The connection is AG-UI, so LangGraph, CrewAI, Mastra, Pydantic AI, Google ADK and a plain HTTP agent you wrote yourself all work. Your model and orchestration layer stay where they are.

A channel is a transport and a rendering target, not a second agent. Messages are authored once as JSX and lowered into Block Kit in Slack and Adaptive Cards in Teams. Generative UI, human-in-the-loop approvals, tools, persistence, slash commands and mentions all ship. @copilotkit/channels 0.5.0 is live on npm under MIT; deployment needs Node.js 22+, an ESM project, and a long-running process.

Channels SDK on GitHub → · AG-UI protocol · CopilotKit

04 · Genspark · Office suite

Genspark Open Sources GenOffice: A Free, Ad-Free AI Office Suite for macOS and Windows with Docs, Sheets, Slides, PDF

Every alternative office suite breaks a Word document eventually, because it re-serializes the whole file on save. GenOffice archives the original by hash and never touches it. Only dirty blocks become OOXML fragments and get spliced back into the original document.xml. Untouched blocks keep their original bytes, and every other zip entry is copied verbatim.

Five Electron apps over one engine layer, Apache-2.0, signed installers for Apple Silicon and Windows x64 at version 0.4.110. Sheets pairs the Univer core with a Rust xlsx sidecar using calamine and IronCalc; Slides and the docx engine are in-house. The team says one engineer, one week, roughly $10,000 in tokens for the Alpha — and Alpha is the operative word: AI runs through a signed-in Genspark account and consumes credits, with no local model key by default.

Read on Marktechpost →  · GitHub  · Download  · Security notes

05 · Marktechpost · RAG tutorial

Pixel-Native RAG: A Practical Guide to Visual Document Indexing

A parser is per-site glue code. A renderer is one code path for HTML, PDF, scanned fax and spreadsheet export alike. This build skips text extraction entirely: render pages as images, slice them into overlapping tiles, embed with SigLIP, CLIP or an optional Qwen3-VL backend, and search a FAISS index.

Tiles are 1024×1024 with 128px of vertical overlap, so a sentence or table row is not split across two embeddings. OCR-based BM25 runs alongside dense retrieval and the two are combined with reciprocal rank fusion; tile scores are max-pooled per document. The tutorial measures Recall@1, @3, @5 and MRR, trains a residual contrastive adapter on mined pseudo-queries, and serves the whole thing behind a FastAPI POST /search.

Read on Marktechpost →  · Full codes  · PixelRAG upstream

06 · NVIDIA · Skill security

Building an Advanced AI Skill Security Auditing Pipeline with NVIDIA SkillSpector, LangGraph, YARA Rules, SARIF, and CI Policy Gates

Agent skills are code you install from strangers. This walkthrough builds a synthetic skill marketplace with clean, risky, malicious and MCP-based examples, then scans each one through SkillSpector’s LangGraph inspection pipeline and keeps the evidence for every finding.

It covers risk scores, categorized findings, confidence levels and analyzer completeness; SARIF and Markdown export; baseline suppressions that survive a deliberately introduced regression; organization-specific YARA rules; a custom secret analyzer added as an extra LangGraph node; and a CI gate that blocks on score, severity, confidence and rule id. The malicious sample exercises prompt injection, env harvesting, credential access and a staged exec chain. Needs Python >=3.12.

Read on Marktechpost →  · Full codes  · SkillSpector

07 · Moonshot · Eval tutorial

Evaluating Multimodal Vision Models with Moonshot PerceptionBench Using Robust Data Loading and Automated Judging

One overall accuracy score tells you a model is worse. It never tells you where. PerceptionBench measures fine-grained visual perception across OCR, counting, localization, contextual reasoning, comparison, depth understanding and hallucination detection, and this harness reports each capability separately.

The pipeline handles resilient multi-stage dataset streaming, base64 image decoding and interleaved image placeholders, then runs three backends: a blind-prior baseline that needs no API key or GPU, OpenAI-compatible multimodal APIs, and local Hugging Face vision-language models. Judging is rule-based with optional LLM assistance, and results carry bootstrap confidence intervals broken out by capability, multi-image questions, image resolution and answer format.

Read on Marktechpost →  · Tutorial codes

Sponsored · Datalab

Marker v2: documents in, Markdown out — 76.0 on olmOCR-bench at 2.9 pages per second

The story above about keeping the original applies to your ingestion layer too. Marker takes PDFs, scans, DOCX and PPTX and returns Markdown, JSON or HTML with tables, inline math and reading order intact — not a flattened text blob with the structure stripped out.

Datalab rewrote Marker as a three-mode pipeline. Version 2 scores 76.0 on olmOCR-bench and sustains 2.9 pages per second on a single B200 — over 5x MinerU’s pipeline backend, and ahead of Docling on both accuracy and speed. Apache-2.0, fully open source.

datalab-to/marker → · Marker v2 vs MinerU, Docling, LiteParse

One more thing…

Y Combinator · Agent harness

Y Combinator Open-Sources QM: An MIT-Licensed Multiplayer Agent Harness That Runs In Slack And The Web

Most agents are designed as personal assistants. Stretch one assistant across a whole company and it gets complicated fast. QM does the opposite. Every person and every room gets its own scoped memory, files, keychain view, permissions, crons, web apps and durable sandbox. The same identity carries between Slack and the web app.

Every turn runs through a headless core over Postgres, and the agent has a small fixed tool surface with one execute tool that runs inside the scope’s own sandbox. Pi, OpenCode, Codex and Claude Code all drive the same core, so a deployment is not tied to one vendor. Three org security postures gate tool calls — Strict pauses every harness tool call, Auto screens external data with a classifier, Dangerous removes both — and hard denials for recursive deletes and destructive SQL apply in all three.

Here is the part that makes it worth the closer. YC runs QM across accounting, legal, events and engineering, including building QM itself — and still calls the release early and buggy. That is the honest version of a dogfooding claim.

Read on Marktechpost → · Project page · Deployment · Threat model

Sponsor AI Dev Signals

Put your product in front of the people who build with it.

This list is engineers, researchers, and founders who read a release note before they read a press release. No banner farms, no interstitials. One sponsor per issue, written in your voice, placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.

See one in action — this issue’s placement for Mend.io →