Ali Asif

The AI log, agents and coding · open weights · science · robotics

llama.cpp v1.0.0, Antigravity /teamwork, GLM-5.3 on two DGX Stations

llama.cpp and ggml hit 1.0, Google shows multi-agent proof results, NVIDIA serves a 753B model locally, and Ai2 tests whether models catch bad premises.

By , published

What shipped

  1. ggml-org

    llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem

    llama.cpp and ggml both reach v1.0.0, with semantic versioning so breaking changes to the public C API now show up as major-version bumps.

  2. Google Research

    Advancing science and engineering with Antigravity's /teamwork

    Google reports multi-agent results in proofs, Lean formalization and NMR structure solving, and opens /teamwork and /boost to Antigravity users.

  3. NVIDIA

    Serving GLM-5.3 across two NVIDIA DGX Stations with vLLM

    A step-by-step recipe for running the 753B-parameter GLM-5.3 NVFP4 checkpoint on two GB300 desk machines, with measured throughput under load.

  4. Allen AI

    XYBench: Helping with the problem behind the question

    An 8,115-question benchmark showing five models tend to answer the mistaken approach a user asked about instead of the goal behind it.

  5. NVIDIA

    Adapt Human Motion for Robots with NVIDIA SOMA Retargeter

    An open source tool that turns human motion data into joint trajectories for five humanoid robots, with reference configs in the v0.2 release.

Five primary pages I opened today. Two are about the plumbing for running models yourself, one is a big multi-agent claim from Google, one is a benchmark about a failure I hit often, and one is robotics tooling. Every number below comes from those pages. I did not run any of it.

llama.cpp and ggml reach v1.0.0

The ggml-org team released v1.0.0 of both llama.cpp (libllama) and ggml (libggml). The point is the contract. From now on, a breaking change to the public C API in the include folder bumps the major version, backward-compatible features and model support bump the minor, and bug fixes bump the patch. The team says more downstream apps depend on the library every month, so it will be stricter about marking those boundaries.

The post also lists distribution channels: Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for iOS and macOS, plus pre-release artifacts for Linux, macOS, Windows, Android and iOS. It says llama.cpp supports more than 150 architecture variants and all mainstream speculative decoding methods. CI runs on self-hosted runners with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.

Antigravity /teamwork

Google Research describes /teamwork as a multi-agent system that pairs parallel exploration with falsifiers such as the Lean 4 proof checker, unmodified test suites and simulators. Each agent gets only the files and tools its step needs, and critics review results from a clean context.

The reported results are wide. Researchers from Google, CUHK, USC and CMU used its Long Proof pattern on the Courtade-Kumar conjecture and formalized the proof in Lean 4 at 50 million lines. A separate /teamwork-LEAP run produced a machine-checked formalization of the Weil conjectures, about 1.8 million lines of Lean across 28,800 subagent invocations. For molecule structure from NMR data, 30 agents reached 95% top-1 exact match on benchmark molecules. Google also says 578 STOC 2027 submissions opted into its Deep Review feedback tool. The patterns are available in Antigravity now.

GLM-5.3 on two DGX Stations

NVIDIA's tutorial serves GLM-5.3 NVFP4, a mixture-of-experts model with 753 billion parameters and 40 billion active per token. The checkpoint is about 464 GB, more than the 252 GB of HBM3e on one GB300 GPU, so two DGX Stations linked over ConnectX-8 RDMA split it with tensor parallelism and expose one OpenAI-compatible endpoint through vLLM. In NVIDIA's AIPerf run on SPEED-Bench Coding, output throughput peaked at 607.8 tokens per second at concurrency 16, then dropped to 385.0 at concurrency 32.

XYBench

Ai2 built XYBench from 8,115 questions drawn from Stack Overflow, Stack Exchange, curated examples and WikiHow. Each asks for help with an approach (Y) when the real goal (X) needs something else, like thickening frosting when the cupcakes are still warm. On Stack Overflow questions, the five models tested focused mainly on the underlying goal in only 35 to 41% of answers, against 55% for the top human answer. The benchmark won Best Resource Paper at EMNLP 2026.

SOMA Retargeter

NVIDIA's open source retargeter converts human motion on the SOMA skeleton into joint trajectories for a specific robot, solving inverse kinematics and exporting CSV. The v0.2 release includes configurations for Unitree G1 and H2, Booster T1, and AgiBot X2 Ultra and A3T3.

What I'd watch next

Whether llama.cpp's first breaking change after 1.0 lands cleanly as a 2.0. Whether anyone outside Google reproduces a /teamwork proof. And whether the 607.8 tokens per second figure holds for real workloads that are not benchmark prompts.

Sources

Topics: NVIDIA, Google, Ai2, Google Research, GLM, Antigravity, Llama CPP, vLLM