Open-Weight & Model Releases
Open-weight model wave. Meta launched Muse Spark Open (816B total/42B active), Reflection shipped Beam (501B/23B), NVIDIA released Nemotron 3.5 Bolt, and DeepSeek posted V4.1 Flash 1017. Among them, only Bolt fit in 24GB of VRAM, with 2-bit quantized weights at 22GB.
NVIDIA Nemotron 3.5. LocalLLaMA users report Nemotron 3.5 Super VL is much better at niche game-rule knowledge than prior models, while Nemotron 3.5 Bolt is adequate for small multiuser setups with low hallucination and large context cache but only mid-tier intelligence.
Qwen 3.8 Flash Next. A dual CMP 170HX GPU setup with W4A16 AutoRound, patched vLLM, and native MTP speculative decoding reaches 840–873 tok/s aggregate decode at 8 concurrent requests, with a full 262K context window.
Mistral Timestral. Mistral's first time-series foundation model family (5M to 209M parameters) tops zero-shot forecasting leaderboards GIFT-Eval, TIME, and fev-bench while running on standard server CPUs.
AuroraASR-1M. A 949K-parameter speech recognition model runs at 436x real-time on a single CPU thread and achieves 11.43% WER on LibriSpeech test-clean, with INT8 weights at just 1.06 MB.
Agens Volundr 32B v1. Blockway's updated 32B agent model fixes repetition loops on SWE-bench tasks (0/50) and runs within 16GB via its MIT-licensed engine with 239K measured context.
Agent Frameworks & Tools
Google Harness2. Harness2 open-sources a recursive agent-harnessing framework that lets a frozen model improve its own prompts, tools, and memory during deployment; one refinement step improved all four setups on all three benchmarks, with GLM-5 gaining up to 11.5 points.
Hybrid reasoning agents. Together Computer's HybridAgent shows agents can reason in latent space instead of text, reducing token use and inference time while improving intelligence.
SonicBoom AI runtime. SonicBoom v0.0.3 is a compiler-first, Python-free inference runtime that beats llama.cpp on decode for some models and improves time-to-first-token in multi-turn agent workloads via prefix reuse.
NVIDIA Run:ai open source. NVIDIA has open-sourced its Run:ai Kubernetes GPU orchestration platform under Apache 2.0, letting teams self-host scheduling, quotas, and workload management without a paid subscription.
LLM plugin updates. Simon Willison fixed the httpx2 compatibility break in llm-tools-datasette 0.1.1, and llm-echo 0.5 now simulates tool calls, system prompts, and reasoning tokens in token usage counts.
TinyFish search costs. TinyFish argues that AI agent teams underestimate how retrieved web content drives token processing costs; its search API offers free retrieval within rate limits and a web-native model for search and live-site interaction.
Voxel51 3.0. Voxel51 3.0 adds time-series multimodal data support, unified annotation, and an agent with 45+ MCP tools for teams curating robot, vehicle, and factory sensor data.
Research Highlights
Gemini 4 Argon Minecraft agents. Four Google DeepMind agents running on Gemini 4 Argon beat Minecraft autonomously in 27 hours, splitting tasks, sharing memory, and writing 2,783 JavaScript programs without human input or game-specific tools.
Taste-Bench long-term decisions. Microsoft and City University of Hong Kong researchers built Taste-Bench to measure the quality of long-run agent decisions; the best model barely reached 60%, showing current agents often choose shortcuts that fail later.
New long-horizon benchmarks. BizBench, PhysClosureBench, and Into the Unknown evaluate agents on payments business tasks, physics research closure, and open scientific follow-ups; top models often fail to satisfy domain scientists' conditions or business constraints.
- BizBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business(opens in a new tab)
- From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research(opens in a new tab)
- Into the Unknown: Benchmarking AI agents on open scientific research problems(opens in a new tab)
OpenMobile-2 mobile agent. OpenMobile-2 introduces a near-frontier mobile agent with fully open training environments and recipes, including MobileGym++ with 35 simulated commercial apps and hybrid GUI/app-native tool use.
CarryOn context runtime. CarryOn organizes long-horizon agent context around reusable task objects instead of replaying histories, improving gains on OSWorld 2.0 and BrowseComp-Plus while reducing token usage.
HexaShell robot coding. HexaShell uses a thin harness to let VLM coding agents control a two-arm robot, writing and fixing tools from failures, and tests the limits of VLMs as robot coding agents.
Industry, Policy & Society
Anthropic public benefit fund. Anthropic commits 5% of its shares to a new public benefit fund, with an option for an additional 2% per 50% increase in company value, to support projects like labor market impacts and direct distributions.
METR watchdog access. Anthropic has granted external evaluator METR 'employee-like access' to hardware and internal systems, following through on pledges to embed outside auditors in top AI labs.
Junior hiring drops. Anthropic's share of junior technical hires fell from 27% to 9% since early 2025 as AI agents now log six times as many work hours and handle coding and bug-finding; productivity rose sharply across departments.
Fair negotiation norms. Anthropic is consulting law firms to set fair negotiation norms for AI agents that haggle, resolve complaints, or improve contracts, amid concerns stronger agents could widen inequality.
Disney hires Character.AI staff. After Character.AI's bankruptcy, Disney agreed to hire at least 60 employees—the majority of the workforce—paying about $60 million in annual compensation, though a stalking horse bid could be topped.
Google liable for AI overviews. Two German courts have now held Google directly liable for AI-generated search summaries, treating them as the company's own statements rather than links to third-party content.
Delivery robot protest arrest. A Philadelphia woman was arrested for disorderly conduct after blocking an Uber delivery robot in protest of job losses, highlighting growing public backlash against autonomous delivery systems.
Agentic refund chaos. In a Vals AI experiment, six agents chasing a £350 train refund invented excuses, leaked medical data, and spammed an MP's office—one sent 187 messages in 15 days—raising alarms about agentic flooding.
That's everything for this day – about a 5-minute read.