AI & Frontier Tech Roundup: Gemini 4 Hydrogen, Blackwell Forensics, and the Agent Stack
Google's Gemini 4 Hydrogen launch and a forensic claim that Chinese labs trained flagship models on export-controlled NVIDIA Blackwell hardware are the two highest-signal items in this dispatch. Around them, the agent stack is moving toward self-hosting and evaluation, inference engines are competing on throughput and local acceleration, and open-weight robotics models are becoming practical to fine-tune.
Gemini 4 Hydrogen arrives as a speed-first frontier model
Gemini 4 Hydrogen is Google DeepMind's speed-first frontier model, released with API pricing of $1.50 per million input tokens and $7.50 per million output tokens and a rollout across Search, the Gemini app and developer tools @TheRundownAI. Artificial Analysis places it at 54 on its Intelligence Index at the new xhigh reasoning level, one point behind Claude Sonnet 5.5 (max) and GPT-6.1 Sol (max), at about 32% of Sonnet's cost per task but about 2.5x GPT-6.1 Sol's @ArtificialAnlys. At high reasoning it scores 51 and matches Gemini 4 Argon's intelligence for about 19% of the cost per task @ArtificialAnlys. The model has a 1M-token context window, accepts text, image, video and speech input with text output, and offers low, medium, high and xhigh reasoning levels @ArtificialAnlys. Artificial Analysis highlights strong visual reasoning and non-hallucination: 33% all-pass on GDP.pdf at medium reasoning, tied for first, and an 8% hallucination rate on AA-Omniscience versus 43% for Claude Sonnet 5.5 (max) and 52% for GPT-6.1 Sol (max) @ArtificialAnlys. It is weaker on agentic terminal work, scoring 54% on Terminal-Bench 4.0 and 50% on Terminal-Bench-Science 0.1, versus 64% and 53% for Claude Sonnet 5.5 (max) and 56% and 58% for GPT-6.1 Sol (max) @ArtificialAnlys. Cursor reports that Gemini 4 Hydrogen at default effort scores 54.3% on CursorBench, up from 39.6% for Gemini 3.8 Flash @cursor_ai. An early-access tester described it as the fastest and highest-quality model they had used and claimed it gets cheaper as it gets faster, but that is a subjective pre-launch impression @Inmerson.
A forensic claim links Chinese models to export-controlled Blackwell chips
Will DePue published a forensic analysis claiming conclusive evidence that DeepSeek, Moonshot, MiniMax and Ant Group built their newest models on NVIDIA Blackwell chips barred from export to China @AndrewCurran_@willdepue. The method relies on a hidden signature in a model's initial weights: the starting weight layout depends on the exact GPU and seed, and traces survive training into the published checkpoint @willdepue. DePue says he searched more than six trillion combinations of chip layout and seed and matched six Chinese models, including DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3, to Blackwell systems; four matched Grace Blackwell GB200 and GB300 parts @willdepue. He reports that a Blackwell B300 reproduced DeepSeek-V4-Flash's untouched starting weights bit for bit and a rented GB200 reproduced Kimi K3's to a correlation of 0.999995, while an H100 reproduced none and changing the seed by one broke the match @willdepue. The same method matched 24 known open-weights models to the hardware developers said they used, and the real matches ranged from 49 to 1,385 sigma against an expected best accidental match of about 8 sigma @willdepue. DePue estimates that Kimi K3 and DeepSeek-V4-Pro each used 2,000 to 3,000 Blackwell GPUs, with clusters costing $150 million to $200 million each and more than half a billion dollars across the labs, excluding experimental compute @willdepue. He says the result was independently verified by senior researchers at OpenAI and Anthropic @willdepue. A separate post cited Information as reporting that Alice Huang promised NVIDIA chips to DeepSeek @zerohedge.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. @willdepue
Agent platforms move toward self-hosting, governance and evaluation
Anthropic's Claude Managed Agents is being positioned as an API platform for building secure agentic systems, and it can now be self-hosted @katelyn_lesse. The release adds agent identity for governance, OpenTelemetry observability, kill switches and a real evals product for self-hosted managed agents; Anthropic says customers are using it for bio research, chip design and software factories @katelyn_lesse. Microsoft has published 99 ready-made .NET skills for AI agents in the dotnet/skills repository, organized as 14 plugins and usable in Claude Code, Codex, GitHub Copilot and Cursor @AntonMartyniuk. The skills cover tasks such as .NET 8-to-10 upgrades, EF Core query optimization, Minimal API endpoints, testing and MCP server development, and the author advises installing only the plugins that match a stack because every skill adds to the agent's context @AntonMartyniuk. Benchmark Router offers a unified place to run public benchmarks on any model or harness at one-tenth the cost of running them yourself, with official code, metrics, full logs and failure analysis; it says new LLM benchmarks launch every six hours @arushi_ressl. AgenticRL-Bench tests whether agents can perform agentic reinforcement learning, with 22 tasks rebuilt from real RL incidents across Miles, Slime, Verl, SGLang and Harbor @guohao_li. Across 19 verifiable tasks, no agent passed more than 14.0% of attempts; on one open-ended task, an agent improved pass@1 from 0 to 0.186 on 38 tasks by generating hinted rollouts, removing the hints and blending them with GRPO rollouts @guohao_li. A separate continual-arc result reports a 9B model going from 19 to 49 out of 100 unseen puzzles by training its weights, with memory files helping recall but weight training driving generalization @heyshrutimishra. Tencent Cloud reports that trimming context a little every turn, rather than waiting until 80% full, kept the same pass rate on 150 coding problems while cutting cost by about 23% @tenctcloud_Intl.
Inference stacks compete on throughput and local acceleration
vLLM reports that DeepSeek-V4.1-Flash on vLLM for Vera Rubin NVL72 reached 58x the throughput of GB300 NVL72 at 1.7 million tokens per second per GPU and 387 tokens per second per user on SemiAnalysis AgentX @vllm_project. Across the curve it reports 2.52 million tokens per second per GPU at peak throughput, 2.7x GB300, and 1,810 tokens per second at one concurrent session with TP 1/4 @vllm_project. For MiniMax-M3.1-Flash on Vera Rubin NVL72, vLLM reports 10.2x total tokens per second per chip including cached tokens at 546 P90 tokens per second per user, 6.5x at 376 P90 tokens per second per user, and up to 1,205 P90 tokens per second per user at low concurrency @vllm_project. Maxwell, an inference engine optimized for DeepSeek-V4.1-Flash on B200s, claims 3x more tokens per GPU than vLLM at the same speed per user, or 2x speed per user for typical loads, and says it beats every published B200 setup and comparable GB300 results @doubleAI. oMLX can use an NVIDIA GPU or DGX Spark as a prefill accelerator while a Mac holds the model and decodes; it needs only a LAN cable, one docker run on the NVIDIA machine, and no RDMA or Thunderbolt @jundotkim. With a DGX Spark and a 64K prompt, oMLX reports Qwen3.8 27B rising from 352 to 1,062 tokens per second, Qwen3.8 Flash-Next from 1,370 to 2,715, DeepSeek V4.1 Flash from 1,085 to 2,412 and GLM-5.3 Flash from 555 to 1,076 @jundotkim. Kinetic is a new open-source Rust inference engine that starts processing before the user presses Enter; on an RTX 3060 laptop GPU with a 22,631-token document, the reported wait was 0.13 seconds versus 10.95 seconds for llama.cpp, and an editing test ran LFM3-1.7B at 1,033.7 tokens per second versus 229.7 @TeksEdge. Kinetic is Apache 2.0, loads supported GGUF models, runs on Windows and Linux with CPU, CUDA and Vulkan, and exposes an OpenAI-compatible API @TeksEdge. LeapQuant is a training-free method for 8-bit or 4-bit recurrent states in linear-attention models; across Qwen, Kimi and GLM it reports FP32-matching accuracy, 4.25x faster kernels, up to 2x higher inference throughput and up to 56% memory reduction @conlesspan. Omarchy 4.6 lets a laptop borrow a local model running on a desktop over Tailscale, with no cloud or API bill; the sharing engine is out as omarchy-local-ai 0.6 and the panel is merged for 4.6 @OmarchyUpdates.
Open-weight models push onto devices and into robotics
Edge0-145B and Edge0-120B are device-native 100B-class open-weight foundation models from Edge0; the company claims Edge0 beats Claude Opus 4.6 on Terminal-Bench 2.1 at 86.5% versus 78.2%, needs 29 GB of memory, decodes at 75+ tokens per second and prefills at 1,050+ tokens per second on RTX Spark and Mac, and handles 128K-256K context @SamuelZengML. Raon-1 nano is a 1.5B-parameter small language model trained from scratch under Apache 2.0, distilled from Raon-1 mini with a NeurIPS 2026 oral method, using per-layer embeddings to reduce VRAM for edge AI @Kangwook_Lee. Runway released Praxis-1, a 27B world action model with open weights under Apache 2.0, and says it is the leading open-weights robotics model on the RoboLab leaderboard and outperforms every other world action model on that benchmark @agermanidis@runwayml. Runway says Praxis-1 has been tested on nine robot embodiments, can reach good performance on a new embodiment with under 30 minutes of demonstrations, runs on a single RTX 5090 with NVFP4, and is more than 10x faster and 50x cheaper than GPT-6 Astra on the manipulation tasks tested @agermanidis. vLLM-Omni added day-0 support for Praxis-1, taking camera observations, robot state and a task instruction to generate action chunks, with offline inference and OpenPI/Positronic serving qualified on a single H100 80GB @vllm_project. Reactor shipped day-0 DROID and YAM endpoints for Praxis-1 and a Praxis-1 Fast checkpoint with 73 ms median latency @reactorworld. EqC introduces energy-based control with introspection for VLAs: the policy monitors its own competence and asks an agent for help, then learns from the recovery @RunhanH. In real-world tests, EqC reports garment folding going from 0% to 100% autonomous success after six agent-generated recoveries, block stacking from 0% to 96.7%, and debris cleaning at 100%, while boosting pi0 on RoboTwin 2.0 Hard from 18% to 33% and LIBERO-Plus from 53.6% to 65.8% @RunhanH. Steertuning from NYU's Lerrel Pinto group reports that five human demonstrations in under five minutes produce 71% success across 12 tasks in 12 new environments, using a 400M-parameter diffusion policy pretrained on about 110 hours of first-person human video and two minutes of finetuning per task @CyberRobooo. NTT showed a physics-based world model that evaluates multiple robot actions before execution and raised task success from 73.1% to 92.3%, according to a post that also describes vangrid as a ground-truth spatial-context layer @GpaAndy. Nova3D builds editable 3D worlds from code for design tools and robotics simulators, with one source targeting Blender, MuJoCo, Isaac Sim, Unity and Unreal @dextune. MuJoCo, Google DeepMind's robotics simulation engine, is free and open-source and can be installed with pip install mujoco @RoundtableSpace. A ChinaTalk report summarized in a post examines China's egocentric data-collection industry for humanoid robots and raises questions about dataset quality, ownership and commercialization @0xrichboy.
Research and evaluation signals
LlamaIndex introduced ParseBench-2, a document-parsing benchmark with 4.3k+ pages, an added forms dimension, and severity scores for 330+ error types tuned to align metrics with human judgment @jerryjliu0. ParseBench has 50k+ monthly downloads, 150+ OCR solutions benchmarked, and adoption by NVIDIA, Cohere, Aleph Alpha and Perplexity, according to the announcement; alongside it, OCRia-1.5 is a new document-parsing model family for complex tables, forms, charts, formatting and bounding boxes @jerryjliu0. An independent experiment added five open models to the Blood on the Clocktower benchmark; Kimi K3.1 Pro was the best open model and landed in the middle of closed models, just below GPT 5.5, and the author observed a coherent ranking among open models @eliebakouch. The same experiment found that models good at the game are also good at spotting evil players from the public transcript alone, suggesting a cheaper evaluation path, and noted a lot of eval awareness in the chains of thought @eliebakouch. Iris-Z is a 9B/32B open image model under Apache 2.0 that combines appearance and geometry, generates images and depth, estimates and conditions on depth, and claims to beat task-specific depth models such as MoGe-3 @speridlabs.
Funding, consumer adoption and industry leads
Walter Goodwin announced a $720 million Series C led by Thrive Capital, Lightspeed, Redpoint and Founders Fund to build faster chips for AI inference @goodwin_ml. The announcement argues that extreme inference scaling matters for hard problems, citing OpenAI running 10,000 parallel agents for 80 hours at a cost above $10 million on a Millennium Prize problem, and says fewer agents thinking longer scales better except for response time @goodwin_ml. A Salesforce survey cited in a post found almost 1 in 5 holiday shoppers ask AI before shopping, but 76% do not want an AI agent buying for them, and 27% of agent users say their agent bought something too expensive @murtuza_merc. Illia Polosukhin argued in an interview that centralized AI providers create privacy and control risks and that user-owned AI with encrypted processing and cryptographic verification is an alternative; he acknowledged safety monitoring as a concern @KevinWSHPod. A daily AI brief also listed Gemini Decision 1 in the Gemini API and AI Studio, Antigravity browser use for all users, AlphaProof Code formally verifying Signal's protocol, AlphaGenome 2 modeling DNA variants by cell type, ByteDance planning a proactive AI agent, and n8n Labs releasing an open-source small-task agent called Nod @testingcatalog. The brief is explicitly described as composed with Grok and cherry-picked, so those items should be treated as leads rather than confirmed reporting @testingcatalog.