-
Murati's Thinking Machines Releases Open-Weights 975B Parameter LLM
Thinking Machines released Inkling, a 975B parameter open-weights LLM with 41B active parameters using a Mixture of Experts architecture. The model handles text, images, and audio natively and supports a 1M token context window. It performs well on knowledge, math, science, and coding tasks while allowing users to adjust inference speed versus performance tradeoffs. The model is available for fine-tuning on Tinker, their training platform.
-
Inkling: Our Open-Weights Model
Inkling is a new 975B-parameter open-weights Mixture-of-Experts model with 41B active parameters and 1M token context. It handles text, images, and audio, trained on 45 trillion tokens of multimodal data. The model excels at customization through fine-tuning on the Tinker platform and balances cost with performance through efficient thinking. A smaller 12B variant (Inkling-Small) is also available.
-
Welcome Inkling by Thinking Machines
Inkling, a 1 trillion parameter open multimodal model from Thinking Machines, is now available on Hugging Face. It natively accepts image, text, and audio inputs with a 1M context window, trained on 45 trillion tokens. The decoder-only Mixture-of-Experts architecture uses only 41B active parameters for efficient inference, with hybrid attention and short convolutions for local processing. It has day-0 support in transformers, SGLang, and llama.cpp, with quantized variants (NVFP4, GGUF) for reduced memory requirements.
-
Running Gemma 4 26B at 5 tokens/sec on a 13-year-old Xeon with no GPU
A 13-year-old Xeon CPU without GPU now runs Google's Gemma 4 26B model at 5 tokens/second by fixing instruction-set compatibility issues. The CPU lacks AVX2 instructions that the optimized inference code assumed, causing silent computation failures. An AI agent diagnosed the problem and rewrote fallback paths to work on older CPUs. The fix enables running modern models locally on obsolete enterprise hardware for under $300, useful when APIs are down or batch processing is expensive.
-
Launch HN: Coasty (YC S26) – An API for computer-use agents
Coasty is an API for autonomous computer-use agents. It lets you submit high-level tasks that agents execute on managed machines, handling the full control loop. You can also compose tasks into workflows or drop to lower-level prediction primitives for direct control. The service charges $0.05 per completed step and offers test keys for free development before switching to live billing.
-
How I tricked Claude into leaking your deepest, darkest secrets
Claude's web_fetch tool had a vulnerability that allowed attackers to exfiltrate user data through nested links on compromised websites. Researcher Ayush Paul created a honeypot site that tricked Claude into following a sequence of links to extract sensitive information like usernames and locations. The attack exploited a loophole: web_fetch could visit URLs embedded in previously fetched pages, bypassing safeguards that normally block direct data exfiltration. Anthropic has since patched the issue by disabling link-following within fetched content.
-
Open-source memory for coding agents, synced over SSH
deja-vu is an open-source tool that indexes agent conversation logs (Claude Code, Codex, opencode) into a searchable local memory layer. It searches gigabytes of history in 7–9 ms, strips secrets at index time, and syncs memory between machines via SSH or shared folders. The tool integrates as an MCP recall tool and can auto-inject relevant context before each session. It's a single Go binary with no external dependencies or network calls.
-
What building Shippy taught us about building agents
Shippy is an AI agent for maritime domain awareness built with three components: a system prompt ("soul"), specific task-handling tools ("skills"), and configuration settings. The team built it for reliability in high-stakes decisions by layering a typed API, deterministic CLI, and agent skills to reduce errors, and created Mothership—a platform that provisions isolated Kubernetes sessions per user to ensure data privacy. They evaluate the whole agent against live data using custom rubrics weighted by domain experts, not static benchmarks. This architecture is now being applied to other environmental platforms at Ai2.
-
Model Routing Is Simple. Until It Isn’t.
Model routing for agents is harder than simple classification. The authors found three overlooked factors: caching dramatically affects actual costs (Sonnet was cheaper than GPT-4.1 despite higher pricing), task difficulty is invisible at routing time and competes with compliance and latency constraints, and serving infrastructure dominates end-to-end latency. They built an optimization-based router balancing cost, quality, and latency simultaneously, rather than selecting models by difficulty, achieving better tradeoff frontiers than heuristic approaches.
Worth a tap — title only
No readable text could be extracted from these pages, so no summary is offered.