Meta Superintelligence Labs ships Muse Glimmer, its first open model
Meta Superintelligence Labs released Muse Glimmer, its first open model — a 30-billion-parameter dense, multimodal model distilled from the larger Muse Spark and tuned for local agentic workflows: function calling, code generation, document analysis, and LLM-as-judge evaluation. It's released under Apache 2.0, and quantized to roughly 4-bit precision it fits in about 20GB of RAM/VRAM, running on a single consumer GPU or a high-end Mac. Day-one support is already live across Ollama, Hugging Face's transformers, and Unsloth's dynamic quants, with llama.cpp, MLX, and ExecuTorch integrations "launching shortly." Meta's own benchmarks put it competitive with similarly-sized open models — Gemma4-31B and Qwen3.6-27B — on agentic, coding, and reasoning tests; that's a vendor claim, not yet independently verified. The release lands the same week Mark Zuckerberg publicly criticized "closed" AI rivals, underscoring Meta's return to an open-weights strategy after Llama's earlier stumbles — and it's a distinct model from the Muse Code terminal coding agent Meta shipped five days ago on updated Muse Spark 1.2.
Sources & depth
Nvidia lines up six financial giants to bankroll $500B of AI compute
Nvidia announced non-binding agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build independent financing platforms aimed at mobilizing over $500 billion in third-party capital for AI compute infrastructure. The pitch, from CEO Jensen Huang, is to turn Nvidia-powered data centers into an investable asset class of their own — "AI factories" that generate revenue like any other productive infrastructure, backed by arguments about GPU longevity, workload flexibility, and resale value across customers. This is a set of memoranda of understanding, not signed deals: no specific allocations, rates, or per-partner commitments were disclosed, and everything is contingent on final agreements. It's the clearest signal yet that the next bottleneck in the AI buildout is financing capacity, not chip supply.
Sources & depth
vLLM v0.27.0 ships full-stack Kimi K3 support
vLLM shipped v0.27.0, a major release spanning 561 commits from 242 contributors, headlined by full-stack support for Moonshot's Kimi K3 — core model and kernel code, Python and Rust frontends, DeepGEMM integration, and compressed-tensors quantized checkpoints all land in a single release. The release also adds FlashAttention 4 integration on newer hardware, a PyTorch 2.13 upgrade, faster DeepSeek-V4 serving, day-one support for Qwen3.5 and several other new models, and a fault-tolerance framework for large-scale distributed deployments. It's one of the more consequential vLLM releases this year for anyone serving open-weight models at scale.
Sources & depth
- github.comv0.27.02026-08-11 02:48 IST
Needle2: a 14MB agentic LLM for microcontrollers and $200 phones
Cactus Compute released Needle2, a follow-up to its earlier Needle model: a 45M-parameter agentic LLM compressed to a single 14MB binary via quantization-aware training (not post-hoc compression), aimed at tool calling and structured extraction on genuinely resource-constrained hardware — microcontrollers, wearables, and sub-$200 phones, which the team notes make up roughly four in five edge devices. It runs a full session in 28MB of RAM, decodes at 500+ tokens/sec on a Raspberry Pi 5, and uses a bounded 256-token sliding window plus grammar-constrained decoding to keep memory and output format predictable with no GPU. It's Apache 2.0-licensed and ships as a dependency-free C++ binary — a niche but genuine data point on how small a useful tool-calling model can get.
Sources & depth