A collection of 336 posts
Cohere released Command A+ on May 20, 2026: a 218B sparse Mixture-of-Experts model with 25B active parameters, Apache 2.0 licensed, that runs on as few as 2 H100 GPUs. Built for sovereign, on-prem enterprise agents with native citations.
xAI's Grok 4.3 lands with a 1M token context window, native video input, and aggressive pricing at $1.25 input / $2.50 output per million tokens. Here is what changed from Grok 4.20, how it benchmarks against Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro, and when it is the right tool to reach for.
Anthropic Mythos is the frontier preview model unveiled April 7, 2026: stronger than Opus 4.7 on math and security, withheld from public release, shipped only via Project Glasswing to ~50 defensive-security partners.
Claude Mythos, Opus 4.7, and GPT-5.5 shipped within three weeks of each other in April 2026. We break down which frontier model wins on coding, reasoning, vision, cost, and which one your team should actually pick.
Alibaba's Qwen 3.7 Max launched May 20, 2026 with a 1M-token context, native extended-thinking mode, and benchmark wins on SWE-Pro and Terminal-Bench. Here's how it compares to Claude Opus 4.7, GPT-5.5, Gemini 3.5 Flash and DeepSeek V4, what it costs on DashScope, and when to pick it.
Google dropped Gemini 3.5 Flash and Gemini Spark at I/O 2026. A frontier-grade Flash model that outruns 3.1 Pro, and a persistent personal agent built on top of it. Here's what shipped, what's rumored, and where it fits next to Claude Opus 4.7 and GPT-5.5.
A practitioner's roundup of every AI model release that mattered in May 2026 — Anthropic Mythos, Gemini 3.5 Flash, Qwen 3.7 Max, Mistral Medium 3.5, ERNIE 5.1, and Subquadratic's 12M-token SubQ. Benchmarks, pricing, availability, and what to actually use.
Baidu's ERNIE 5.1, released May 8 2026, became the first Chinese LLM in the global Search Arena top 5. Here's what it does, how it compares to DeepSeek V4 and Qwen, and when teams outside China should actually use it.
Mistral Medium 3.5 is a 128B dense model with built-in reasoning, coding, and agentic capabilities. Le Chat Work Mode turns it into a multi-tool agent. Here's what's new, what it costs, and when to actually pick Mistral over Claude or GPT.
Manus AI in May 2026: the Meta $2B acquisition was blocked by China's state planner on April 27, the Desktop app's My Computer feature puts the agent on your local machine, and pricing, access, and the autonomous-agent landscape sit where they sit.
Subquadratic, a Miami startup, launched SubQ on May 5, 2026 — the first frontier LLM with a 12-million-token context window, built on a new Subquadratic Selective Attention (SSA) architecture. Here's what's verified, what's plausible, and what to actually do with it.
DeepWiki turns any public GitHub repo into an AI-generated wiki at deepwiki.com. The URL swap, Fast vs Deep Research, private repos via Devin, MCP for Cursor and Claude Code, vs GitHub and Cursor.
What LM Studio is, how to install it on Mac, Windows and Linux, how the OpenAI-compatible server works, MLX vs llama.cpp on Apple Silicon, document chat (RAG), the lms CLI, and where it beats Ollama and llama.cpp.
What a 4x RTX Pro 6000 Blackwell rig actually buys you for DeepSeek V4 Flash: throughput, VRAM headroom, the SM120 vLLM compile bug, and break-even vs the DeepSeek API.
An honest 2026 guide to llms.txt: what the spec actually says, what adoption looks like in server logs (the SERanking 300k-domain study), real annotated examples from Stripe and Anthropic, the robots.txt + AI-bot User-Agent stack that actually works, and a copy-pasteable template.
Kimi K2.6 ties GPT-5.5 on SWE-bench Pro at 58.6% — and runs roughly 3x cheaper, with open weights. Where each model wins, with the cost math.
Kimi K2.6 ties Opus 4.7 on multilingual SWE-bench but trails by 7 points on Verified — at 1/5th the cost. The honest, benchmark-by-benchmark breakdown.
Kimi K2.6 and DeepSeek V4 Pro are the two best open-weights coding models in 2026. K2.6 wins long-horizon agents and swarms; DeepSeek V4 wins on raw price.
DeepSeek V4 Flash is the under-covered story of the V4 release. 1M context, 47 on the AA Intelligence Index, $0.14 input / $0.28 output per million tokens, and it fits on a Mac Studio. Here is the full practical guide.
A deep, engineer-focused comparison of DeepSeek V4 Pro vs DeepSeek V4 Flash: benchmarks, pricing, speed, local deployment, and a decision tree for picking the right variant for your workload in 2026.
Eight days apart, Anthropic and DeepSeek shipped the two most consequential AI releases of 2026. Here is the honest, benchmark-backed comparison engineering leaders need before they re-architect their stack.
DeepSeek V4 launched the same week as GPT-5.5 and GPT-5.5 Pro. We break down the benchmarks, pricing, 1M-context engineering, coding wins, and which model your team should actually deploy.
Quick answer. The DeepSeek V4 API is OpenAI-compatible. Point any OpenAI SDK at https://api.deepseek.com, set your key, and call deepseek-v4-pro (top reasoning/agentic) or deepseek-v4-flash (cheap, fast). Minimal request: curl https://api.deepseek.com/chat/completions -H "Authorization: Bearer $DEEPSEEK_API_KEY" -H "Content-Type:
Quick answer. For pure SWE-bench Pro top score and 1M-context agentic coding, pick Claude Opus 4.7. For longest-horizon swarm runs, pick Kimi K2.6 — open-weight and roughly 8x cheaper. For broad reasoning + Codex/CLI tooling, GPT-5.5. For commodity-priced inference at frontier-adjacent quality, DeepSeek V4 Pro. Choose per workload,
Learn how to run MiniMax‑M2.7 locally using GGUF, llama.cpp, and vLLM, with hardware needs, benchmarks, pricing, and examples.