Meta open-sources 816B Muse Spark Open weights

The MI355X's 288GB of HBM3E per GPU gives it an edge on very large models like the 1.6T-parameter DeepSeek V4 Pro, but serving-stack maturity on ROCm remains the bottleneck at higher loads. Teams evaluating accelerator purchases for agentic inference can now compare AMD and NVIDIA on realistic coding-agent workloads rather than synthetic throughput numbers. All results run on ROCm nightly builds of vLLM and SGLang, so scores should improve as the software stack matures.

官方账号Artificial Analysis@ArtificialAnlysSource

Crossing US$1B ARR while raising at US$15.75B signals that serverless GPU infrastructure has become a core revenue layer for AI workloads, competing with hyperscalers on consumption-based pricing. Engineering teams building inference or training pipelines should check Modal's GPU Glossary, a four-part reference from CUDA architecture to profiling that doubles as a crash course in GPU performance engineering. Its pay-per-second model suits startups in any region, including Singapore, that lack dedicated GPU capacity.

shao__meng@shao__mengSource

The interview reframes the compute gap as the main constraint on Chinese labs rather than talent, and treats distillation as a months-long edge instead of an existential leak. Builders evaluating open-weight strategies get a concrete map of the RL-environment market and the reward-hacking pitfalls in purchased environments. Trillium's plan to give away environments for free directly undercuts paid vendors like those selling proprietary RL tooling.

a16z@a16zSource

The course is structured around three engineering pillars—decomposition, data, and evaluation—giving practitioners a curriculum-level map of what production agent work actually demands beyond prompt tinkering. Developers and ML engineers building agents can use the free lecture notes as a self-study path, especially the coverage of DSPy, MCP and memory architectures. Its critical reading list, including 'Why Do Multi-Agent LLM Systems Fail?', offers a rare sceptical counterweight to agent hype.

shao__meng@shao__mengSource

The lesson is that 98% GPU utilisation can hide a CPU-bound training loop, and profilers — not utilisation dashboards — are the right diagnostic tool. Engineers fine-tuning linear-attention or hybrid architectures like Qwen's gated delta net can often get large speedups from existing PyPI kernels rather than rewrites. Those avoiding FSDP2 dynamics should also note why CUDA Graphs and torch.compile were ruled out here.

shao__meng@shao__mengSource

Bitwise consistency between rollout and training removes a silent source of RL instability that most teams discover only as unexplained reward degradation. Engineers running RL fine-tuning loops on vLLM can now keep optimized inference paths without sacrificing training fidelity. It is a practical reference for anyone comparing inference engines for post-training pipelines.

官方账号vLLM@vllm_projectSource