News archive
Browse the archive
- One issue per day at a permanent URL.
- Reverse-chronological order; most recent first.
- Per-layer item counts on each card show which layers moved that day.
- Click any card to open the full issue, or subscribe via RSS.
- 2026-05-25 safety-guardrails 2 · training 1 · weights 1
- Today's issue covers items from 2026-05-22 through 2026-05-25 across three layers: training, weights, and safety-guardrails.
- NVIDIA released Nemotron-Labs Diffusion, a family of open-weight diffusion language models at 3B, 8B, and 14B plus an 8B vision-language model, under the NVIDIA Nemotron Open Model License with weights on Hugging Face.
- The training layer saw TRL 1.5.0 add chat-template and async-GRPO support, and the safety-guardrails layer logged two items: NeMo Guardrails 0.22.0 and Anthropic's first Project Glasswing update on AI-assisted vulnerability discovery.
- 2026-05-21 weights 2 · safety-guardrails 2 · runtime 1 · agents 1 · governance 1
- Today's issue covers items from 2026-05-19, 2026-05-20, and 2026-05-21 across five layers: weights, runtime, agents, safety-guardrails, and governance.
- The weights layer recorded two open releases, with Cohere's Command A+ as the most material primary release: a 218B-parameter mixture-of-experts model with 25B active parameters, released under Apache 2.0 on Hugging Face.
- The runtime layer logged llama.cpp build b9254, which adds Programmatic Dependent Launch support for NVIDIA Hopper and newer GPUs.
- The safety-guardrails layer logged two items, including Prime Intellect's controlled reward-hacking findings and a related Prime Sprints credit-and-prize program.
- 2026-05-19 governance 2 · runtime 1 · retrieval-memory 1 · agents 1 · protocols 1
- Today's issue covers items from 2026-05-18 and 2026-05-19 across five layers: runtime, retrieval-memory, agents, protocols, and governance.
- The runtime layer recorded three same-day Ollama v0.30.0 release candidates (rc18, rc19, rc20) on 2026-05-18, continuing the rebase away from GGML toward direct llama.cpp and MLX support.
- The most significant primary release is HuggingFace's Ettin Reranker family, six Apache-2.0 ModernBERT cross-encoders from 17M to 1B parameters, with the 32M model reportedly outscoring BAAI bge-reranker-v2-m3 (568M) on MTEB Retrieval NDCG@10.
- The protocols layer logged Anthropic's acquisition of Stainless, the SDK and MCP-server generator that has powered Anthropic's official SDKs since the API's earliest days.
- 2026-05-18 runtime 4 · weights 2 · evaluation 1 · governance 1
- This issue covers items from 2026-05-15 through 2026-05-18 across four layers.
- The runtime layer carried four primary-source items: vLLM v0.21.0 stable, SGLang v0.5.12, Ollama v0.24.0 stable, and Ollama v0.30.0-rc17.
- The weights layer logged Nathan Lambert's Latest Open Artifacts #21 on Interconnects and IBM's Granite Embedding Multilingual R2.
- The evaluation layer carried IBM Research's Open Agent Leaderboard announcement, and the governance layer logged Jack Clark's Import AI 457.
- 2026-05-14 runtime 3 · agents 2 · evaluation 1
- Today the runtime layer carried three primary-source items: vLLM v0.21.0rc3, Ollama v0.24.0-rc0, and a Hugging Face write-up on asynchronous continuous batching that reports a 22% generation-time speedup on an 8B model.
- The agents layer logged Anthropic's $200M four-year Gates Foundation partnership and Latent Space's coverage of competitive shifts between OpenAI Codex and Claude API metering.
- The evaluation layer carried Allen AI's launch of AIMIP, a climate-model intercomparison effort joined by NVIDIA, Google Research, and several universities.
- The most significant primary release of the day is the Hugging Face transformers asynchronous-batching contribution, which delivers a measurable inference-throughput gain without new kernels.
- 2026-05-13 runtime 3 · silicon 2 · governance 2 · weights 1
- This is the inaugural issue of the daily roundup.
- Today the runtime layer carried three primary-source releases (vLLM v0.21.0rc2 and v0.20.2 plus SGLang v0.5.11), the silicon layer logged a SemiAnalysis write-up on Cerebras wafer-scale inference, and the governance layer covered both the OSI's new executive director update and Jack Clark's Import AI 456.
- The most significant primary release of the day is vLLM v0.21.0rc2, which packages bundled DeepGEMM and wires the nvidia-cutlass-dsl[cu13] extra for CUDA 13 platforms.