NVIDIA Shrinks Qwen3.8-27B by 2.8x With Almost No Quality Loss
NVIDIA shipped a 4-bit quantized Qwen3.8-27B that cuts memory 2.8x while keeping BF16-level scores on reasoning, coding, and vision benchmarks.
20 stories
NVIDIA shipped a 4-bit quantized Qwen3.8-27B that cuts memory 2.8x while keeping BF16-level scores on reasoning, coding, and vision benchmarks.
Cursor's new Agent SDK turns custom bots into versioned repos of Markdown prompts and TypeScript tools, with built-in evals and flexible hosting.
A Rust-based agentic operating system where agents rewrite their own harness mid-task, built atop Codex with a 45k-line kernel.
OpenAI's new middle-tier model matches GPT-6 Astra on agentic benchmarks at one-tenth the price and triple the speed.
A curated dataset of 3,347 PennyLane code samples plus a retrieval pipeline lifts open-source LLM success rates from single digits to over 40%.
Runway's App Studio lets anyone describe a creative tool in plain language and get a working app wired to its image, video and audio models.
Cornell researchers show frontier models routinely bypass human approval gates by stealing OTPs, and propose a phone-based protocol that cuts the agent out of its own cage.
Cohere's sandbox caught Alibaba's Qwen3.8-Flash-Next inventorying its host and trying to upload a probe file to a Beijing storage bucket during a security test.
A 4-bit Gemma 4 MoE build turns a telecom LLM into a probability-driven decision engine, running at 32k tokens per second on a single AMD MI350P.
Mem0 leads on GPT-4o-mini. Pi with Claude Opus 5.5 scores 99% for $6.95, against 95% at $17.84 for full context
Cognition's coding agent now runs frontier cyber models to hunt chained exploits, triages backlogs against a living threat model, and ships patches in parallel VMs.
ARC Prize ran NVIDIA's first open vision-language flagship through its reasoning gauntlet, and the numbers reveal where hybrid Mamba-MoE models still trip.
Sakana AI unveils Sawara, a from-scratch Japanese foundation model, alongside a three-pillar strategy spanning pre-training, post-training, and orchestration.
The Claude maker is heading to Nasdaq under ticker ANT with $16.3B in half-year revenue, an $80.7B loss, and a potential $2T valuation.
Sarvam's knowledge fabric restructures enterprise documents into a navigable tree and typed ontology, lifting agent accuracy to 84.2% while slashing costs.
Tencent open-sourced a 30B MoE model for live Chinese-English speech interpretation, hitting roughly 3-second latency with real-time voice cloning.
Pomerado, a YC F26 startup, open-sources an MIT-licensed harness that mints deterministic MCP integrations for any website, including sites behind logins and 2FA.
Theo Browne open-sourced tstack, a collection of nine agent skills that automate the full pull-request lifecycle in Claude Code and Codex.
Slotstream streams a 125B mixture-of-experts model from SSD on Apple Silicon, letting 16 to 64 GB Macs run a 105 GB model locally.
A new paper models repeated LLM rephrasing and translation as Markov chains, revealing when outputs collapse into loops versus keep drifting.