Local AI guides
Run local AI on your own hardware - practical, cost-driven guides.
Arete-27B: a Qwen retrain that drops state framing
Nous fine-tuned Qwen3.8-27B with 52.8M tokens of supervised data. Answers carrying state propaganda fell from 22% to 0.9% and capability benchmarks stayed flat.
Nov 07, 2026
Nvidia's Spark laptop runs Linux at one third speed
An encrypted power table locks the RTX Spark laptop's GPU to 780 MHz on native Linux: 13.9 tok/s against 65.4 on Windows. What it means for Spark buyers.
Nov 05, 2026
Memory squeeze: why every local AI build costs more
Apple cut its margin guidance on memory costs, the RTX 5090 hit $9,400, and SK hynix eyes $100B of NAND fabs. What it does to every local build.
Nov 05, 2026
Kimi K3.1 Pro: 2.8T open weights beat Opus 5.5 on agents
Kimi K3.1 Pro ships 2.8T open weights with 104B active and 1M context. It beats Opus 5.5 on agentic search and automation, and costs 1.56TB to host.
Nov 05, 2026
Ripple-65B: a dense 65B model that beats 500B open weights
Ripple-65B Preview scores 80.52 on TerminalBench 2.1, ahead of open-weight models up to 975B, and runs at 33GB in 4-bit under Apache 2.0.
Nov 05, 2026
Orca: 115k stars for the harness that runs your agents
Orca runs five coding agents at once, each in its own git worktree, on your own subscriptions. MIT, 115,729 stars, SSH worktrees for your rig.
Nov 05, 2026
Strata: a 125B model on a 12GB card, measured
Strata runs a 125B MoE model on 12GB of VRAM. The measured numbers say 4 to 11 tok/s, and the DeepSeek claim needs 191GB to 274GB on disk.
Nov 05, 2026
Superintelligence ban: the bill Congress could actually pass
H.R. 10538 bans superintelligence and pauses advanced AI training above 10^25 FLOP. What it does to open weights, charters, distillation and home rigs.
Nov 05, 2026
DeepSeek V4.1 Flash: the subagent tier gets an open model
DeepSeek V4.1 Flash cut KV cache to 890 bytes per token and now wins three of four agent benchmarks against Opus 5. Here is the routing split teams are testing.
Nov 05, 2026
No AI bailouts: Congress opens the bankruptcy file
House Democrats' October 28 resolution says failing AI firms go through ordinary bankruptcy. Inside the $518B to $807B obligation stack that prompted it.
Oct 29, 2026
GitHub Universe: Copilot became the agent harness
GitHub shipped Fusion, multi-agent projects, and local model routing inside one desktop harness. What Universe 2026 changed and who it serves.
Oct 29, 2026
Anthropic's S-1: $80.7B loss, 95% accounting marks
Anthropic S-1: $16.3B H1 revenue (1,300% growth), an $80.7B loss that is 95% noncash note marks, $693M operating loss, $417B compute. Nasdaq: ANT.
Oct 27, 2026