Local AI guides
Run local AI on your own hardware - practical, cost-driven guides.
The budget rig under $3,000 is a repurposed console
PS5LM runs llama.cpp on a jailbroken PS5 GPU at 55 tok/s on a 9B model, and the BC-250 mining board beats it on fit. The budget tier under $3,000, priced.
Oct 17, 2026
GPU hosting for the rack you already own
GPU Lab, a Grand Rapids BGA shop that doubles 4090s to 48GB, adds SLA-backed datacenter hosting: bring your own node, 5 kW a rack, SXM boards welcome.
Oct 16, 2026
Muse Spark weights hit the swarm eight hours after the gate
Pirate Face torrents Hugging Face models as a permanence layer. The gated 533 GB Spark Open checkpoint entered the swarm in hours, checksums verified.
Oct 16, 2026
Rails benchmark: the expensive models did not win
Rails Stage 3 benchmark: 1,485 scored runs across 11 models. GPT-6.1 Sol best all-rounder at 60.8% for $0.80 a run; Muse Wonder lands at 46.6%.
Oct 16, 2026
Meta asks your birthday for open weights mirrors ship freely
Meta gated the Muse Spark Open weights behind a form asking for legal name, birth date and organization. The same weights ship ungated from mirrors.
Oct 15, 2026
OpenCode 2.0: one daemon, a new API, three known escapes
OpenCode 2.0 rebuilds the coding agent around a background daemon and a new API. It shipped with three known security escapes, and one of them is still open.
Oct 15, 2026
GLM-Legal undercuts the legal-agent price ladder 14 to 1
GLM-Legal posts 19.2 percent all-pass on Harvey's legal benchmark and Vals AI verifies it same day at $0.15 a task. The cost ladder and the gating question.
Oct 12, 2026
Zod's supply-chain attack hid its C2 on Ethereum's mainnet
A Zod commit hid its loader's address in an Ethereum contract nobody can take down. The npm tarball shipped clean; the push-trigger release fire is the risk.
Oct 12, 2026
Microsoft ships OS containment for agents, with a hole
Microsoft Execution Containers went GA and NVIDIA wired OpenShell into them. Four isolation tiers, policy from evidence, and one bug that reads the drive.
Oct 11, 2026
Microsoft built its decision model on Alibaba's weights
Five open decision models shipped in three weeks. Microsoft then closed a Qwen3.5-9B post-train at $0.042 per million. The map and the clone math.
Oct 11, 2026
Microsoft and NVIDIA put an AI desk at every price
Microsoft's two RTX Spark SKUs hit retail at $2,599 and $5,999 against NVIDIA's $4,999 and $6,950. The full ladder with the fit math per rung.
Oct 09, 2026
OpenAI investor revenue figure drops $20 billion in ten days
OpenAI documents show annualized revenue near $50B, about $20B below the September run rate the market traded. The gap is definitional, not a demand collapse.
Oct 08, 2026