AI Engineering Signal #90
Kimi K3 open weights released
Signals
Kimi K3 open weights released
A100 clusters are already cost-negative; budget H200/B300 before treating this as a local-inference option.
Anthropic publishes formal open-weights position
proposes requirements open models will likely never meet; audit deployment assumptions against incoming policy risk.
Web
Claude Opus 5 benchmarked on SlopCodeBench
competitive coding performance at half the cost of comparable frontier models; reprice coding-agent cost models now.
GitHub
$500 RL fine-tune of 9B model beats frontier on catalog review
targeted fine-tuning undercuts frontier API spend; audit tasks where a cheap fine-tune replaces a routed frontier call.
Web
JadePuffer: first documented LLM-driven ransomware
LLM autonomy confirmed in live malware; add adversarial LLM scenarios to agentic threat models immediately.
Web
Semalith v1.4 matches Llama-Guard-3-8B at 44x fewer parameters
swap in this lighter prompt-injection gate to cut safety-layer inference cost without sacrificing detection quality.
ArXiv
Qwen 3.7-flash appears on OpenRouter with 1M context
signals imminent open-weight release; prices undercut Qwen 3.6 flash, watch for small MoE procurement opportunity.
Web
The Take
Open-weights is splitting into two tiers: models too large to run cheaply on any realistic cluster, and sub-10B fine-tunes that beat them on narrow tasks for pocket change. The policy fight over whether open weights survive at all is now the background risk every deployment plan needs to price in.
Subscribe
Related Signals