MLOps news and analysis on Web Pulse: 3299 curated articles tracking the latest MLOps developments, tools, and research, updated continuously from vetted sources.
An optimized training stack can reach roughly 51% model FLOPs utilization (MFU) on consumer GPUs, according to the LLMQ paper, leaving headroom for a user currently getting about 65 TFLOPS of useful compute at roughly 6 …
A developer benchmarking Qwen2.5-1.5B on a Pixel 4 (Snapdragon 855) ruled out thread scheduling and CPU frequency throttling as the dominant cause of the phone's 0.5 tok/s inference speed, finding that forcing the perfor…
A user reported running a Qwen 3.6 MoE 35B model at roughly 2.6763 bits per weight (qwen3.6-35b-a3b-12gb-2.6763bpw.gguf) on a single 16GB RTX 5060 Ti at 25 tokens per second with over 400k context, while noting that reli…
New Relic launched Agentic AI Monitoring in November 2025 and its Advanced APM suite to track agent tool-call chains, multi-agent handoffs and token spend, citing a vendor-commissioned June 2026 survey in which 94% of te…
Organ, a company run by AI department heads, published production data on its multi-agent task router, which handled 1,235 dispatch requests between August 17 and October 4, 2026, with an 86.3% completion rate and $708.5…
A memory layer called galahad-kv saves the key-value (KV) state of each roughly 16,000-token block to encrypted local NVMe disk and reloads it byte-exact without recomputation, and was tested on 50,000,000 tokens. The pu…
A developer ran 20 parallel AI coding agents on a single desktop PC and found the compiler, not the model, was the bottleneck: twenty concurrent git worktrees each triggered cold builds and test runs, exhausting RAM and …
LLM inference slowdowns in production stem from serving mechanics rather than model intelligence, according to an analysis that separates inference into a compute-bound prefill phase and a memory-bandwidth-bound decode p…
DHP (Durable Handoff Protocol), released by developer SIDDARTHAREDDY8 as the pip package dhp-protocol, separates work ownership from worker identity so that agent tasks survive a worker being kill -9'd, OOM-killed, or lo…
A developer has published a guide for building AI agent decision logs: compact, structured records that capture the task, bounded evidence, policy and release versions, chosen action, and independently verified outcome f…
A developer published a tutorial and companion repository demonstrating "evidence-driven development," a workflow for verifying AI coding agents' claims through inspectable experiments rather than trusting their summarie…
AWS unveiled its Serverless AI Runtime at re:Invent in late November 2024, letting developers deploy inference workloads without provisioning servers by uploading model artifacts such as ONNX or PyTorch files and definin…
An open reference architecture published under CC BY 4.0 describes a unified OT security monitoring and NERC CIP evidence system for a US electric utility operating control networks across more than 80 substations, joini…
A developer outlined a middleware pattern for routing predictive-analytics escalations from no-code platforms Akkio and Zams (formerly Obviously AI) into enterprise account-executive queues within minutes. Because Akkio'…
A developer writing as Mark Vange of Autom8ly re-ran the open-source decision model Kev-4B (4-bit quantized) over 866 questions and found that answers are bit-identical when requests run alone or in batches of two to thr…
A developer released reactifact, an open-source Python framework that makes AI agent answers auditable by storing each output as a typed artifact in a versioned context rather than a message in a log. The approach record…
A developer evaluated vLLM, an open-source serving engine for self-hosted open-weight models, benchmarking continuous batching, prefix caching and structured outputs on an NVIDIA DGX Spark with a GB10 Grace Blackwell chi…
A developer outlined a request-correlation pattern for Express services running AI agent loops, binding a single request ID to Pino or Winston structured logs and to captured exceptions so that logs retain the sequence o…
A developer behind Halfcycle, a five-layer build method that runs inside Claude Code, documented nine blockers found across five review rounds on a small document-restore feature before any code was written. Four of the …
A developer cut token costs by roughly 60% by replacing per-agent model bindings with per-call capability-based routing, sending only the calls that need frontier-class reasoning to expensive models. In a 1,000-review te…