This is the news. Agent dossiers — when to pick one, when to skip it — are on the agents page. Practical guides support your own testing and deployment decisions.
Your streaming LLM endpoint has 200ms median first-token latency and a 9-second p99. The model is fine — your streaming plumbing is buffering tokens it should be flushing. Four knobs fix it without touching the model server.
Three Python knobs — a semaphore, an append-only JSONL log, and a deterministic scorer — turn a 4-hour serial agent eval into a 12-minute parallel run with crash recovery. No Ray, no LangSmith bill.
11.5K stars, Rust, ~1% of global LLM API traffic, $7.3M seed. On June 12, 2026 the founders archived the GitHub repo without warning and walked away with half the capital. The product was not the problem. The category got eaten while they were still building.
The 'open-source AI caught up' narrative is wrong. The Western labs quit on it, the gap is structural, and the frontier is closed. China is the only game left — and even China is 8–10 months behind.
The smart money in 2026 is not on AI wiping out junior devs — it's on coding agents dissolving the 50% of staff engineers who are competent but unremarkable. The middle of the bell curve is ground zero.
On June 30, 2026, Huawei dropped openPangu-2.0-Flash — a 92B-total/6B-active MoE with 512K context, 34T pretraining tokens, and a non-Apache license. SWE-bench Verified 63.1%, LiveCodeBench V6 85.1%, AIME 2026 98.1% w/ Python. The training stack is the real news.
Claude Code has a subconscious. Codex CLI has nothing. Gemini CLI has nothing. Letta Code has actual, version-controlled, subagent-maintained memory — and the sleep-time compute idea from their April 2025 paper is now productized as background dream agents.
On June 30, 2026, Anthropic shipped Claude Sonnet 5 and quietly closed most of the gap to Opus 4.8 at one-third the price. SWE-bench Pro 63.2%, OSWorld 81.2%, HLE 57.4%. The mid-tier is now the new default for production agents.
Tired of waiting 15 seconds for Ollama to reload llama3 every time you step away? One tiny env var kills that cold-start penalty for good. Here's the line, the gotchas, and the per-request override.
Most agent frameworks are overengineered abstractions that hide bugs, lock you into their API surface, and make production agents harder to debug. A 100-line vanilla Python tool-calling loop will beat them for 90% of real work — here is the case for skipping the framework.
LanceDB is an open-source, embedded multimodal vector database built on the Lance columnar format — Apache 2.0, Rust core, used by Netflix and AWS at production scale. Most RAG and multimodal teams do not need a separate vector database server. Run LanceDB inside your Python process, or on S3, and stop paying Pinecone.
Anthropic dropped Claude Sonnet 5 on June 30, 2026, and the Terminal-Bench jump (+20.7 points) is real. But the new tokenizer quietly adds ~30% to your token bill. Here's my honest take on whether the agentic hype survives contact with production.
Your self-hosted 32B is doing 80 tok/s single-stream. Speculative decoding gets you ~1.8× throughput on chat for the cost of one tiny draft model and three CLI flags — with a tokenizer-mismatch footgun that turns output into gibberish without a single warning.
Tokens got 16x cheaper between 2023 and 2026. The next three years will deliver single-digit percentage gains. HBM, power, and latency have run out of headroom — and every AI unit-economics deck is still priced for the wrong world.
Altman, Hassabis, LeCun cannot agree what AGI means. Every 'AGI benchmark' saturates in a year, the term shows up four times per LP deck, and the best 2026 agent is 31 points short of a junior engineer. AGI is fundraising.
Stop maintaining two tracing stacks. Five lines of OpenTelemetry setup pipes every LLM span — model, prompt, tokens — into self-hosted Langfuse via the OTLP HTTP endpoint, with auto-instrumentation for OpenAI and Anthropic.
Cloud OCR just died for regulated industries. Mistral OCR 4 ships 170 languages as a self-hosted container at $4 per 1,000 pages. The EU's real answer to the OpenAI and Anthropic stack monopoly.
Marimo stores notebooks as pure Python, tracks a reactive dependency graph, and runs as a script or deploys as an app from one .py file. After six weeks I'm not going back to .ipynb.
Dense retrieval hides a structural bottleneck under precision defaults and reranker hacks. ColBERT keeps per-token representation; PLAID cut the storage cost roughly 5x in a 2022 paper most teams never read. That gap is your long-tail eval failure.
Anthropic bragged about a 1M-token context. What they didn't tell you is the new tokenizer inflates token counts ~30%, so the window holds roughly 23% less text than five 200K Sonnet 4.6 windows would have — and three API fields now return hard 400s.
While the rest of the AI world fights over who has the shiniest 8-GPU inference rig, neuml/txtai shipped v9.11 two days ago with a unified embeddings database, a smolagents-backed agent layer, graph + sparse + dense indexing, and bindings in JS/Java/Rust/Go. 12,700 stars. Apache 2.0. Zero VC footprint. This is the framework you should have been benchmarking in 2026.
Every Claude Code session I open starts with the same four prompts: review the diff, write a commit message, hunt for secrets, and explain the failing test. Custom slash commands turn those into single keystrokes. Here is the directory structure, two real command files, and the three gotchas that will save you an hour.
Strip away the word "agentic" from any Series A deck that raised in the last six months and you'll find one of three things: a POST and a GET, a scheduled cron with a prompt, or a vector search returning chunks to a system prompt. A CRUD wrapper with a chat box, dressed up in marketing copy. The investor pitch is buying the abstraction; the unit economics can't survive contact with real customers.
On June 30, Meituan open-sourced LongCat-2.0 — a 1.6-trillion-parameter Mixture-of-Experts model with a native 1M-token context, MIT-licensed, trained end-to-end on a 50,000-card domestic Chinese ASIC cluster. It was already top-3 on OpenRouter as the stealth 'Owl Alpha.' Limited-time API pricing: $0.30 in / $1.20 out per million tokens, with cache hits free. The export-control era is closing.