Model Releases

  • deepseek-ai/DeepSeek-V4-Pro-DSpark — DeepSeek V4 Pro released in DSpark format with FP8 quantization support and endpoints compatibility. Accompanied by arXiv paper 2606.19348, this is a significant update to the DeepSeek model family with 8‑bit FP8 precision. 🤖📄
  • nvidia/GLM-5.2-NVFP4 — NVIDIA’s Model Optimizer quantized version of GLM‑5.2 to 4‑bit FP4 precision, showcasing inference optimization on the MoE architecture. MIT license. 🔥
  • krea/Krea-2-Turbo — Krea‑2‑Turbo is a fast text‑to‑image diffusion model built on Krea‑2‑Raw with a custom Krea2Pipeline for Diffusers integration. Optimized inference variant for speed. 🛠️

Open Source Releases

  • vllm-cpu-nightly 0.23.1.dev202606280855 — Nightly build of vLLM’s CPU‑targeted inference engine, focused on high‑throughput and memory‑efficient LLM serving without GPU. Good for CPU‑only deployments. 📦
  • kvquant-plus-plus 0.1.12 — Attention‑aware KV cache quantization library extending KVQuant++ to reduce memory footprint during autoregressive generation. Helpful for long‑context inference. 🧠
  • vllm-sr 0.3.0.dev20260628084824 — vLLM Semantic Router for intelligent routing across Mixture‑of‑Models setups, enabling dynamic model selection based on query semantics. Useful for multi‑model serving. 🛠️
  • tfp-nightly 0.26.0.dev20260628 — Nightly build of TensorFlow Probability, providing probabilistic modeling tools integrated with TensorFlow. For Bayesian and uncertainty work. 📊
  • OpenSpec: Spec-driven development for AI coding assistants — OpenSpec introduces a spec‑driven development approach for AI coding assistants, guiding code generation with formal specs instead of ad‑hoc prompting. Relevant for structured AI coding. 📝
  • pygeovision 2.0.7 — Geospatial AI platform with a 119‑model registry (DINOv3, Prithvi‑EO‑2.0), FastAPI serving, edge/cloud deployment, 51 pipelines and 25 notebooks. For Earth observation and remote sensing. 🌍

AI Dev Tools

Today’s Synthesis

Deploying DeepSeek‑V4‑Pro‑DSpark on a CPU‑only stack just got realistic: pull the FP8‑quantized model from Hugging Face, spin up the vllm-cpu‑nightly build (0.23.1.dev202606280855) and let it load the weights directly into CPU memory. The real kicker is kvquant‑plusplus (0.1.12), which slices the KV cache with attention‑aware quantization—perfect for the model’s MoE layers and long‑context generation. In practice, you’ll configure vLLM’s kv_cache_dtype="fp8" and enable the kvquant‑plusplus plugin, then launch inference with --max-model-len 8k and --gpu-memory-utilization 0.9 (the latter is a no‑op on CPU but keeps the CLI happy). The result: a ~30 % drop in CPU RAM usage while preserving FP8 precision, letting you run the 8‑bit DeepSeek model on a modest server without GPU acceleration. 🤖🧠📦