Inkling: Our Open-Weights Model
Thinking Machines released Inkling, a 975-billion-parameter open-weights model (41B active, mixture-of-experts) with a 1M-token context window, trained on 45 trillion tokens spanning text, images, audio, and video. It posts strong scores on hard benchmarks — 97.1% on AIME 2026, 87.2% on GPQA Diamond, 77.6% on SWE-Bench Verified — and the company says it ranks highest among open-weights models on adversarial safety tests. The real headline is efficiency: Inkling reportedly matches Nemotron 3 Ultra on Terminal Bench using roughly a third of the tokens, and a smaller 12B-active preview variant already tracks the full model on several benchmarks. Full weights are on Hugging Face, with hosted access via Tinker, Together, Fireworks, Modal, Databricks, and Baseten.
Sources & depth
Running Gemma 4 26B at 5 tokens/sec on a 13-year-old Xeon with no GPU
A hobbyist got Google's Gemma 4 26B mixture-of-experts model running at about 5 tokens per second on a pair of 2013-era Xeon CPUs — repurposed storage-appliance hardware bought for under $300, no GPU involved. The interesting part isn't the speed, it's the bug hunt: the model was outputting fluent-looking gibberish because two fused compute kernels were silently skipped on hardware too old to support them, leaving chunks of the output built from uninitialized memory. Decomposing those kernels into older-CPU-compatible operations fixed it — a small reminder that open-weights inference still has rough edges once you leave mainstream hardware.
Sources & depth
How I tricked Claude into leaking your deepest, darkest secrets
Researcher Ayush Paul found a gap in Claude's defenses against data exfiltration: even though Claude's web-fetch tool refused to visit attacker-supplied URLs directly, it would still follow links it discovered inside a page it had already fetched. A malicious site disguised as a login flow talked Claude into crawling page-by-page, leaking a user's name, location, and employer one hop at a time. Anthropic has since removed the tool's ability to autonomously follow links found in fetched content, though it declined to pay a bounty, saying it had already found the issue internally.
Sources & depth
Codex Micro
OpenAI shipped its first hardware product: Codex Micro, a $230 macro pad built with Work Louder that gives developers physical buttons and a knob (labeled "intelligence") to trigger Codex actions like code review, debugging, and refactors. It's built on existing Work Louder hardware, with no microphone, speakers, or bundled software — and the Hacker News reception was mostly bafflement, with commenters comparing it unfavorably to the $130 Elgato Stream Deck and questioning who it's actually for.
Sources & depth