Meta shakes up the frontiers, open and closedInterconnects5 minKIForschung
Meta's Muse Spark Open becomes the strongest US open model, and with Reflection's Beam, US open weights are starting to close the frontier gap with Chinese models.
So · Tech-Tagesreport · Ausgabe 98
2026-10-18
— Open models are ballooning to 552B parameters, yet security incidents remind us: the foundation still isn't stable.
TL;DR des Tages
DeepSeek releases V4.1-Flash-1017, with across-the-board gains on reasoning, coding and agent benchmarks while cutting output length and cost, setting a new bar for open-weight models. OpenCode suffers a security breach with user credentials leaked, and the team urges key rotation. Perplexity open-sources SPACE, a production-grade Firecracker microVM sandbox for AI agents. StepFun, Meta and others keep pushing in open weights, and US open models are starting to close in on the frontier.
Schlagzeilen
1
DeepSeek V4.1-Flash-1017 released: 552B MoE open model sweeps benchmarksMehrere Quellen ×6
DeepSeek released V4.1-Flash-1017 on October 17, a sparse MoE model built on a Causal Encoder-Decoder (CED) architecture with 552B total parameters, 8B prefill activation and 16B decode activation, supporting 1M context and native image understanding. Versus the September version, DeepSWE v1.1 rose from 74.2 to 77.0 and Terminal-Bench 4.0 from 31.2 to 42.3, while task execution time and output length both fell. Why it matters: The model is released as open weights under the MIT license and served via the deepseek-flash API, with Search and Fetch APIs also opened up, giving coding agents and long-running tasks a new option on the performance/cost Pareto frontier and directly challenging closed-model pricing logic in agent scenarios.
Commenters broadly acknowledge its strong performance and high cost-effectiveness, but some question the benchmark results, arguing real-world performance may not beat closed models like Sol.
2
OpenCode breached, user credentials leaked, officials urge key rotation
OpenCode confirmed a security breach on the night of October 17 involving a large-scale distillation attack that led to user credential leaks. An official notice advises all users to rotate API keys immediately. Why it matters: OpenCode is an AI coding tool for developers, and leaked credentials mean attackers could directly call users' cloud service APIs, causing financial loss or data leaks. For engineers relying on the tool to access LLMs, this is a real-world warning about supply-chain security.
Commenters broadly believe the OpenCode Console leaking user API keys had been warned about and the team is to blame, though some argue only the Console service is affected and local CLI users are untouched.
3
Perplexity open-sources SPACE: an AI agent sandbox built on Firecracker microVMs
Perplexity has open-sourced SPACE, the production-grade sandbox platform behind every Perplexity Computer task. Each agent task runs in an isolated Firecracker microVM, with credentials and outbound traffic controlled from outside the VM, and the code is released on GitHub under the Apache 2.0 license. Why it matters: For teams building coding or browsing agents, SPACE offers a production-validated self-hosted isolation layer. Its 0/108 VM escape interception record and 60ms median creation latency provide a directly reusable infrastructure reference for secure agent execution.
4
StepFun releases Step 5 Preview open weights: 600B MoE, 27B active, 1M context
StepFun released the Step 5 Preview open-weight model with 600B total parameters, 27B activated per token, and support for up to 1M token context. FP8 and BF16 checkpoints are live on Hugging Face and ModelScope, targeting long-running agentic workloads. Why it matters: This is yet another self-hosted option for coding agents and document analysis. Though Terminal-Bench 4.0 is only 33.3%, it reaches 88.3% on AA-LCR long-context reasoning, and the hardware bar is clear (FP8 needs 4x B300 or 8x H200), offering teams with data-center GPU resources a new deployment choice.
Jeden Morgen ein Tech-Digest, für dich kuratiert
Das Web zeigt das große Ganze; Abonnenten bekommen ihr eigenes — nach deinen Interessen kuratiert, dein privates RSS integriert, mit Community-Stimmen, jeden Morgen zugestellt. Dauerhaft kostenlos.
98 Ausgaben erschienen · täglich 150+ Meldungen auf 30 gesiebt
KI-News
Replit's CEO announces switching the default model from Anthropic to OpenAI, citing cost, partnership and internal benchmark performance.
Research on behavioral composability of sparse autoencoder (SAE) features: features are partially additive, but accurate prediction and steering require interaction terms.
🤖Sparse autoencoder features are partially compositional but require interaction terms for accurate behavioral predictions and steering.
Virtual Head Attentionhf_papers2 minKIForschung
Virtual Head Attention (VHA) uses lightweight linear transforms to simultaneously cut KV cache and training/prefill FLOPs while preserving model quality.
🤖Virtual Head Attention improves long-context model efficiency by reducing KV cache and compute while preserving quality through lightweight linear query and output transformations.
Beyond Policy Rollouts: Embedding-Space Directional Exploration with Anchored Noise for LLM RLVRhf_papers2 minKIForschung
The EDEN method injects anchored embedding noise during the RLVR rollout phase, boosting trajectory diversity and strengthening group-relative advantage signals.
🤖EDEN improves reinforcement learning with verifiable rewards by injecting anchored embedding noise into rollout conditioning to increase diversity and strengthen group-relative advantage signals without altering inference.
Dev & Open Source
Pele renders Mermaid syntax into themeable SVG, just 230 kB gzipped with zero dependencies, rendering two orders of magnitude faster than Mermaid.js.
esp32sim implements instruction-level ESP32 emulation in the browser, able to boot the real mask ROM and unmodified firmware, supporting ESP32-S3/C3/C6.
llmspeed.dev lets users estimate local LLM token generation speed before buying hardware; the community likes the idea but finds estimation accuracy and configuration options limited.
Commenters broadly like the idea of a tool that helps estimate local LLM hardware performance, but some find its speed estimates inaccurate and its model and hardware configuration options limited.
Qwen 27B (IQ4_XS, Q8/Q4 KV) at full 262k context on a 16GB RX 9070 XT: 80+ t/s decode with Adaptive KV Streamingreddit2 minKIChips & HW
With adaptive KV streaming and RDNA 4/ROCm optimizations, Qwen 27B decodes at 80+ t/s on a 16GB RX 9070 XT across the full 262k context.
Community-Themen
2 gaming PCs + an M4 Pro MacBook over Ethernet + 2 iPhone 17 Pro Maxes over USB-C: Qwen3.8 Flash Next at 37-39 tok/s at 60k context (262k layout in the works)reddit3 minKICloud & Infra
Running Qwen3.8 Flash Next distributed across 2 gaming PCs + an M4 Pro MacBook + 2 iPhone 17 Pro Max units over Ethernet/USB-C, hitting 37-39 tok/s at 60k context.
Discussion of how Qwen 3.8 Flash Next's n-gram table (PLE) is over-quantized in common GGUF builds, hurting performance; using the full table is recommended.
A reflection piece by an infrastructure engineer encountering Kubernetes for the first time struck a chord, with comments focusing on criticizing Helm string templating and YAML nesting complexity.
Commenters broadly praise the article's readability and focus on criticizing Helm string templating and layer upon layer of YAML nesting, though some argue the problem isn't YAML or the templating layer, and K8s's value as a desired-state reconciliation engine shouldn't be dismissed because of it.
llama.cpp has come a long wayreddit3 minKIOpen Source
A developer shares the journey from skepticism to acceptance of local AI, marveling at llama.cpp's progress over the past year or so.
GitHub Trending
Star anthropics / claude-plugins-official Official, Anthropic-managed directory of high quality Claude Code Plugins.
MadsLorentzen/ai-job-search★ 50364
Sponsor Star MadsLorentzen / ai-job-search The job search that runs on your machine. AI job application framework built on Claude Code: evaluate postings, tailor CVs, write cover letters, prep interviews. Fork it and own it.
affaan-m/ECC★ 282491
Sponsor Star affaan-m / ECC The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
HKUDS/CLI-Anything★ 54048
Star HKUDS / CLI-Anything "CLI-Anything: Making ALL Software Agent-Native" -- CLI-Hub: https://clianything.cc/
Star K-Dense-AI / scientific-agent-skills Turn any AI agent into an AI Scientist. The #1 Agent Skills library for science, used by 250,000+ scientists worldwide. 177 ready-to-use validated skills plus 100+ scientific databases covering biology, chemistry, medicine, and drug discovery. Compatible with Cursor, Claude Code, Codex, Pi, Antigravity, and the open Agent Skills standard.
akitaonrails/ai-memory★ 9910
Star akitaonrails / ai-memory Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors
android/skills★ 8417
stablyai/orca★ 95390
Star stablyai / orca Orca is the place for working with a fleet of agents (e.g. Grok Bot, Claude Code). Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
anthropics/claude-code★ 151727
Star anthropics / claude-code Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
debpalash/VoiceStudio★ 73312
Sponsor Star debpalash / VoiceStudio VoiceStudio is the open-source, fully-local ElevenLabs alternative — voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages.
Weitere Fundstücke（18 weitere）
Roundtables: A Conversation With the Creator of AI-Designed VirusesMIT Tech Review1 minKIForschung
Listen to the session or watch below Can AI design new life forms? In 2025, Stanford University PhD student Samuel King came up with a preliminary answer when he used a generative AI model to propose genetic blueprints for microscopic viruses. It isn’t yet an example of AI-generated life, but that could be next. Watch…
I’ve been building VeloxDB, an open-source database client built with Rust and Tauri, and it just crossed 650 GitHub stars. The idea came from wanting a database tool that feels fast, lightweight, and comfortable to use without sending everything through a cloud service. VeloxDB supports PostgreSQL, MySQL, SQLite, and more. It runs locally, connects directly to your databases, and includes a visual data explorer and query editor. It’s been a fun journey building this in the open, learning from f
A Meta spokesperson insisted that the updated terms did not reflect a change in the way Meta crawls the web.
After backlash, Meta is updating its terms again to remove language that suggested it could scrape content from sites, including news sites, to train its AI. Specifically, the following language no longer appears on the commercial terms page: … Meta may use any content from such domains, web pages, or URLs to provide, develop, improve, […]
The DNA inside our cells isn’t just a series of instructions, but something far more dynamic. Like scientists, AI models don't yet have the data to properly decode it.
Show HN: Parin – A 2D game engine between raylib and Godot, written in DlangShow HN1 minOpen SourceDev-Tools
Article URL: Comments URL: Points: 38 # Comments: 6
Compression is predictionreddit1 minForschungKI
As video models enable longer and more realistic interactive experiences, maintaining a world consistent with its history becomes essential. This requires both preserving previously observed content and ensuring coherent world evolution over time. World memory is a critical capability to evaluate alongside visual fidelity and video duration. We introduce WorldMemBench, a benchmark for systematically evaluating world memory in video world models. Across 260 test cases, it assesses static memory t
Safebucket v0.8.0 - Self-hosted file sharing, now with customisable links, bulk actions and UI refreshreddit1 minOpen SourceCloud & Infra
Hello! We're two friends working in tech, and we've been building Safebucket: a simple, S3-agnostic, open-source alternative to WeTransfer, Dropbox and Lockself. What's new in v0.8.0: Customisable links: Create a quick share link with a custom slug File explorer improvements: Bulk actions (move, download and trash) Move files over with drag and drop UI/UX redesign We're open to feedback, contributions, and questions. What features would you like to see next? Github: Documentation:
I understand this kind of product may not be exactly right for this subreddit, but I still think it deserves some attention, as it's an LLM frontend for local engines in the first place, and can be used for general-purpose chats with LLMs once lore extraction is disabled in the settings. Otaku is built with a focus on three key ideas: maintain writing quality as the story grows; keep the context small , even for long stories, without losing the details that matter; offer high-quality interfaces