All money pages, every release
All tutorials
-
How to Use Gemini Managed Agents (Background Tasks and Remote MCP in the Interactions API)
Google's July 7, 2026 update to the Gemini Interactions API lets you run a fully managed agent from a single endpoint — Gemini handles reasoning, code execution, package installs, and file management inside an isolated cloud sandbox. The new release adds background execution for long-running tasks (background=true), remote MCP server tools so the agent can reach your private databases and APIs, custom function calling with a requires_action step, and credential refresh that keeps the sandbox warm. This guide shows the verified quickstart code, then each new feature with the exact parameters from Google's official docs.
-
How to Stop llama-server From Throwing Away Your KV Cache (and Make Local Inference Fast Again)
If your local model in llama-server re-processes the entire prompt on every request — or worse, every new chat session — you're paying for prefill you already did. The fix is a handful of official flags most people never turn on: --cache-reuse for KV shifting across shared chunks, --cache-ram and --cache-idle-slots to keep more prompt cache in memory, and --slot-save-path plus the /slots save/restore API for cross-session persistence. This guide explains what llama-server caches by default, why it still discards work (including the SWA gotcha), and the exact flags to fix each case — all from the official llama.cpp server docs.
-
How to Run GLM-5.2 Locally (the MIT 744B Coder, and the 2-bit Build That Fits a 256GB Mac)
GLM-5.2 is Z.ai's flagship open-weight coding and agentic model — MIT-licensed, 744B total / 40B active MoE, a solid 1M-token context, and strong enough to hit 62.1 on SWE-bench Pro and 81.0 on Terminal-Bench 2.1. But 744B parameters means the local route needs serious memory. This guide gives the honest hardware reality, then three real ways to run it: llama.cpp with Unsloth's GGUF quants (the 2-bit build fits a 256GB Mac or a 24GB GPU with offload), the official vLLM / SGLang servers, and the free-to-start Z.ai API — plus the thinking-effort controls and the sampling settings straight from the official model card.
-
How to Run Leanstral 1.5 for Lean 4 Proofs (Local vLLM and the Free API Shortcut)
Leanstral 1.5 is Mistral's Apache-2.0 open-weight model for Lean 4 formal theorem proving and autoformalization — 119B total / 6.5B active MoE, 256K context, and strong enough to saturate miniF2F and solve 587/672 PutnamBench problems. This guide covers what it does, the two real ways to run it — the free Mistral API through the Vibe CLI (`vibe --agent lean`) and a self-hosted vLLM server (4-GPU) — plus the Lean 4 setup, hardware reality, and the recommended sampling settings, all from Mistral's official docs.
-
How to Run Kimi K2.7 Code Locally (and Whether You Actually Should)
Kimi K2.7 Code is Moonshot AI's new 1-trillion-parameter (32B active) open-weight agentic coding model — Modified MIT, 256K context, strong coding benchmarks. But '1T parameters' means the smallest usable local build is ~339 GB of combined RAM+VRAM, not a laptop. This guide gives the honest hardware reality, then three real ways to run it: llama.cpp with Unsloth's GGUF quants (big-memory local), the official vLLM / SGLang servers (datacenter GPUs), and the pragmatic path most people actually want — the Moonshot API or GitHub Copilot, where K2.7 just went GA.
-
How to Switch to Claude Sonnet 5: 3 Breaking Changes and a Hidden 30% Token Cost
Claude Sonnet 5 (model id claude-sonnet-5) is a drop-in upgrade for Sonnet 4.6 — but 'drop-in' hides three API behavior changes that will throw 400 errors if you don't handle them, plus a new tokenizer that makes the same text cost ~30% more tokens. This guide walks the exact migration: the one-line model-ID swap, the three breaking changes (adaptive thinking on by default, sampling parameters rejected, manual extended thinking removed) with before/after code, the token-cost math you need to re-budget, pricing, availability, and when to pick Sonnet 5 over Opus 4.8 — all from Anthropic's official docs.
-
How to Run Ornith-1.0 Locally: The Open-Weight Coding Model That Writes Its Own Scaffold
Ornith-1.0 is DeepReinforce's MIT-licensed agentic-coding model family (9B / 31B dense, 35B / 397B MoE) built on Gemma 4 and Qwen 3.5. The 9B scores 69.4% on SWE-Bench Verified and runs on a normal laptop. This guide shows five ways to run it locally — Ollama and llama.cpp from the official GGUF, LM Studio, Hugging Face transformers (the official quickstart), and an OpenAI-compatible server (vLLM / SGLang) you can point a coding agent at — plus the reasoning-model settings that matter, honest benchmark expectations, and what 'self-scaffolding' actually means.
-
How to Run Apertus Locally: The Fully-Open Swiss LLM (8B & 70B) in 7 Minutes
Apertus is a fully-open, Apache-2.0 LLM from Switzerland's EPFL/ETH Swiss AI Initiative — trained on 15T tokens across 1,800+ languages with every artifact (data scripts, checkpoints, training code) released. This guide shows four ways to run it locally: Ollama, llama.cpp, Hugging Face transformers (the official quickstart), and MLX for Apple Silicon. Includes size/VRAM guidance, the chat template settings that matter, realistic benchmark expectations, and what 'fully open and compliant' actually buys you.
-
GitHub Copilot Usage-Based Billing: How AI Credits Work & How to Control Costs
On June 1, 2026 every GitHub Copilot plan moved to usage-based billing. Premium requests are gone, replaced by GitHub AI Credits charged on token usage. Here's exactly what changed, how many credits each plan includes, what still costs nothing, how to set a budget so you don't get a surprise bill, and what annual subscribers need to know.
-
How to Use MiniMax M3: API, Pricing, Coding Agent & Local Setup
MiniMax M3 is a 428B/23B-active open-weight model with a 1M context and native multimodality, reporting 59.0% on SWE-Bench Pro. Here's how to actually use it: the OpenAI-compatible API, the real pay-as-you-go pricing ($0.30/M input), how to wire it into a coding agent, and what it takes to run locally.
-
How to Build a Batch Document Extraction Pipeline with Qwen3-VL
A production batch pipeline that turns a folder of receipts, invoices, and PDFs into validated JSON and CSV — using Qwen3-VL through an OpenAI-compatible endpoint. Typed Pydantic schema, business-rule validation, retries, and concurrency. Every code block is from a pipeline that really runs; in this guide 3 receipts extract correctly in 5.7 seconds.
-
How to Switch to Claude Fable 5 (API, Claude Code, claude.ai) — and the 3 Things That Change
The model string is the easy part. Claude Fable 5 changes three things Opus 4.8 builders aren't expecting: thinking is always-on adaptive (you can't disable it), refusals come back as HTTP 200, and cyber/bio work is handed to Opus 4.8. Here's the complete switch for every surface, the exact model IDs, and the API gotchas — all from Anthropic's own docs.
-
How to Make Gemma 4 Run up to 2x Faster Locally: Multi-Token Prediction (MTP) + QAT
Two free speedups for local Gemma 4 that people keep confusing. QAT cuts memory ~72%; multi-token prediction (MTP) roughly doubles decode throughput — and MTP just landed in mainline llama.cpp (merged June 7, 2026). The exact Ollama, llama.cpp, and Transformers commands, what speedup to actually expect, and which sizes are supported. Sourced from the merged PR and Google's own docs.
-
How to Run Gemma 4 12B Locally: Ollama, llama.cpp & Transformers (Text, Image, Audio)
Gemma 4 12B is Google's encoder-free open model that runs text, image, and audio on a 16GB laptop under Apache 2.0. The exact Ollama, llama.cpp/GGUF, and Transformers setup — including how to pass images and native audio — all from the official model card and Unsloth docs. Copy-paste safe.
-
How to Run Bonsai Image 4B Locally: On-Device Text-to-Image on Mac & PC
Bonsai Image 4B is a ternary/1-bit diffusion model that fits in ~1 GB and generates a 512×512 image in ~6 s on an M4 Pro — fully on-device, Apache 2.0, zero per-image cost. The exact setup, which quant to download, the one-shot CLI, and the local Studio server, all from PrismML's official demo repo.
-
How to Run NuExtract 3 Locally: vLLM, Templates & Document Extraction
NuExtract 3 is a 4B open-weight VLM that pulls structured JSON out of any document — invoices, receipts, contracts, PDFs — and runs on a single 16GB GPU. The exact vLLM and Transformers setup, the JSON template language, image and multi-page PDF extraction, and document-to-Markdown — all from the official model card, copy-paste safe.
-
Run Qwen3.6-35B-A3B Locally for Coding: llama.cpp, Quants & VRAM
Qwen3.6-35B-A3B activates just 3B params but scores 73.4% on SWE-bench Verified — and it runs on a 24GB GPU or a 32GB Mac. The exact llama.cpp setup, which quant to pick, the VRAM math, and how to wire it into a coding agent.
-
Claude Code as a Daily Driver: CLAUDE.md, Skills, Subagents, Plugins & MCP
Most people use Claude Code like a chat box. The five features that turn it into a daily driver — persistent CLAUDE.md memory, reusable Skills, delegated Subagents, Plugins, and MCP servers — with the exact files, commands, and gotchas for each.
-
Claude Opus 4.8 is live — the 4 changes builders should actually care about (it is not the benchmark numbers)
Anthropic shipped Claude Opus 4.8 on May 28, 2026. Same price as 4.7, modest benchmark gains, but underneath are four shifts that change how you wire up agents: prompt cache minimum drops 4× from 4,096 to 1,024 tokens, fast mode drops 3× cheaper, mid-conversation system messages preserve cache, and Dynamic Workflows runs hundreds of parallel subagents for codebase-scale migrations. Here is what changes for builders.
-
How to Enable Claude Opus 4.8 in Cursor, Claude Code & the API
Claude Opus 4.8 not showing up? It doesn't appear everywhere automatically. Here's how to enable it in Cursor, Claude Code, claude.ai, and the Anthropic API — including the stale-CLI fix when it's missing from your model picker, and the model-ID gotcha that breaks API calls.
-
Anthropic and OpenAI just found product-market fit — the $2,180/month bill Simon Willison made public, and what builders should do about it
Simon Willison's May 27 post says Anthropic and OpenAI have finally reached PMF — through coding agents. The receipts: his own $1,199.79 Claude Code + $980.37 Codex monthly bill, Anthropic's projected $10.9B Q2 revenue, SpaceX's $1.25B/month contract, 25% of Uber's commits from Claude Code. Here is what builders should actually change about their stack in response.
-
Use AI to write better code, more slowly — the multi-agent code review workflow that beats one-shot generation
Nolan Lawson's HN-trending essay nails the inversion: most builders use AI to ship low-quality code fast; the better play is using it to ship high-quality code slowly. Here is his actual workflow, why the EURECOM Constraint Decay benchmark says it works, and why DeepSeek-tier pricing makes it affordable to run 4-7 sub-agents per PR.
-
Constraint Decay — why AI coding agents nail prototypes but break on production backends (and the 8-framework benchmark that proves it)
A new EURECOM paper benchmarks GPT-5.2, Kimi K2.5, MiniMax M2.5, Qwen3-Coder-Next and others across 8 web frameworks and 4 constraint levels. Capable models lose 30 points (40% relative) when you add architecture + database + ORM rules on top of an API spec. Here is what the data actually says, and what to do about it as a builder.
-
Reasonix — the DeepSeek-native coding agent that proves V4 Pro's cache math (99.82% hit, $12 instead of $61)
Reasonix is an MIT-licensed terminal AI coding agent engineered around DeepSeek's prefix cache. The author published a real benchmark — 435M input tokens in a day at 99.82% cache hit, costing $12 instead of $61. Here is what it is, how to install it in 2 minutes, and how it compares to Claude Code / Cursor / Aider.
-
How to add /llms.txt to your site — the LLM-friendly standard, with a real example
Anthropic, Vercel, Cursor, and now 7minai all have a /llms.txt file. It tells LLMs what your site is about and lists your most useful pages — and unlike robots.txt or sitemap.xml, it is markdown that humans can also read. Here is the full spec, a real example you can copy, and the 10-minute Astro implementation.
-
7 Minutes to Master DeepSeek V4 Pro — The 75% Permanent Price Cut Changes the Agent Math
DeepSeek made the V4 Pro 75% price discount permanent on May 22, 2026. Input is now $0.435 / 1M tokens (cache hit: $0.003625) and output $0.87 / 1M — roughly 1/2 of Claude Haiku 4.5, 1/10 of Gemini 3.5 Flash output. Here is exactly what changed, how to use it via API and Chat, the cache-hit math, and when it actually wins.
-
How to Use Cursor Composer 2.5: Setup, Pricing & Benchmarks
How to use Cursor Composer 2.5 in your IDE: step-by-step setup, the real pricing (and the fast-variant trap), and how it compares to Opus 4.7 / GPT-5.5 on coding benchmarks — plus when not to switch.
-
How to Use Qwen3.7-Max: Qwen Chat, Pricing & Benchmarks
How to use Alibaba's Qwen3.7-Max right now: switch to it free in Qwen Chat (no signup), what it actually outputs on a real coding task, the 1,000-tool-call / 35-hour agentic claims, the pricing, and when to pick it over Claude Opus 4.7 or GPT-5.
-
Gemini 3.5 Flash: Pricing, Benchmarks & Whether to Upgrade
Google's agent-tier Gemini 3.5 Flash skips 3.2/3.3/3.4 and lands at $1.50/M in + $9/M out — 3× pricier than 3 Flash but beating 3.1 Pro on most benchmarks. The real pricing, the benchmark wins, and whether the 3× is worth it over the older Flash.
-
50 Best Gemini 3.5 Flash Agent Prompts
Stop using chat prompts for an agent-tier model. 50 prompt templates for Gemini 3.5 Flash, organized by use case, designed to take advantage of the 1M context window and streamed reasoning.
-
Gemini 3.5 Flash vs Claude Haiku 4.5: The Agent-Tier Pick in 2026
Two agent-tier models on paper, two different design philosophies. Compare context windows, pricing, native multimodality, and reasoning behavior — using only published numbers, not made-up benchmarks.
-
7 Minutes to Master Gemini Omni Flash
Gemini Omni Flash is Google's new any-to-video model — drop in text, image, audio, video, or a sketch, get back a 10-second clip with sound. Clips are capped, the API isn't open yet, but pricing leaks at $0.10–0.30 per second. Here's how to use it and how it compares to Veo, Sora, Kling, and Runway.
-
7 Minutes to Master Gemini Spark
Gemini Spark is Google's first proper consumer agent — it runs 24/7 on its own Google Cloud VM, drafts emails by reading your docs, and is powered by Gemini 3.5 + the Antigravity agentic harness. Here's exactly what it does, who can use it today, and how to think about it as a builder.
-
How to Use Gemini 3.5 Flash: Step-by-Step Tutorial
Everything from your first message in the app to setting up the API for agentic workflows. We cover AI Studio, the model ID string for Cursor/Cline, and how to verify you are actually using the 3.5 engine.