Notes from production AI.Not marketing content.
What we learn shipping AI onto client hardware: self-hosting economics, model choices, and security.
OpenCode v1.18.34 has no offline flag. Its source shows 17 outbound paths, four with no setting at all, and the model catalog now calls models.opencode.ai.
Everything we have
written down.
Every post comes out of work we did: a client system, an internal build, or a benchmark we ran. Filter by topic above, or read straight through.
OpenCode Offline Mode: Air-Gapped Setup for Local Models
OpenCode v1.18.34 has no offline flag. Its source shows 17 outbound paths, four with no setting at all, and the model catalog now calls models.opencode.ai.
OCT 02, 2026
vLLM Structured Output Not Working: HTTP 200, No JSON
vLLM removed guided_json in v0.12.0 but still answers it with HTTP 200 and free text. Ten failure modes stamped by version, their fixes, and a CI canary.
OCT 01, 2026
pg_search vs pg_textsearch: BM25 Hybrid Search in Postgres
pg_search vs pg_textsearch: both scored 0.68 nDCG@10 on SciFact, double native ts_rank_cd. Replicas, license and RLS decide which one you ship.
SEP 30, 2026
LeRobot n_action_steps vs chunk_size: How Many to Execute
SmolVLA's own ablation falls from 82.8% to 51.8% on LIBERO when it plays all 50 predicted actions. What n_action_steps and chunk_size do in LeRobot.
SEP 29, 2026
SR 26-2 Model Risk Management: Which LLM Parts Are In Scope
SR 26-2 puts generative and agentic AI out of scope in one footnote. The classifiers, rerankers and scoring layers around your LLM are still models.
SEP 28, 2026
Open WebUI Vulnerabilities: Which Advisories Hit Your Setup
53 Open WebUI advisories landed between July and September 2026, and 30 need a setting that ships off. Map each one to your config, then pin v0.11.4.
SEP 25, 2026
Docling vs MinerU vs Marker (2026): Code and Model Licenses
Marker weights trigger at $5M revenue or $5M funding, MinerU at USD 20M monthly group revenue. Code vs weight licenses for all three, quoted and dated.
SEP 24, 2026
No User Query Found in Messages: Qwen Tool Loop Fix
Qwen 3.5, 3.6 and 3.8 templates reject any request with zero plain user turns. On Ollama qwen3.8, a tool loop outgrowing num_ctx gets there as a 500.
SEP 23, 2026
Mem0 vs Zep vs Letta vs Cognee: Which to Use in 2026
Mem0 publishes 94.4% on LongMemEval, Zep 90.2% at 104ms p50 retrieval. The accuracy argument is over, so pick on tokens, latency and certs.
SEP 22, 2026Not ready to talk?
What we learn shipping AI systems, once a week. No announcements, no fluff.