A collection of 358 posts
GLM-5.3 claims 2,436 real vulnerabilities found — and FreeBSD and Red Hat CVEs actually credit the model line. What verifies, what doesn't, and the licence question that matters most.
GLM-5.3 launched August 14, 2026 with big coding claims and cyber capabilities. The API and independent benchmarks have since arrived — the weights haven't. What shipped, what's verified, and whether to wait.
DeepSeek V4-Pro hit general availability on August 13, 2026. It ranks #2 on SWE-bench Verified at 1/36th the cost — and last of seven on agentic coding. The split is the story.
DeepSeek moved to peak/off-peak billing at 16:00 UTC on August 16, 2026. V4-Pro output rose up to 4.55x and cache-hit input up to 12x. Full rate tables, the peak windows, and how to cut costs under the live pricing.
Alibaba's Qwen3.8-Max is a 2.4T MoE with a near-MIT licence and elite algorithmic coding. But its agentic benchmarks reverse under a neutral harness, and you can't run it locally.
xAI's Grok 4.6 numbers verify cleanly — but the Terminal-Bench version trap, a stale official leaderboard and a 2.3x cost-per-task rise make them easy to misread.
Grok 4.6 improves coding and reasoning but regressed on agentic coding, tripled time-to-first-token and costs 2.3x more per task despite an unchanged price list.
Grok 4.6 launched August 12, 2026. Specs, the 200K pricing cliff, the Terminal-Bench version trap, and an honest read of where it beats and loses to Claude Opus 5 and Muse Spark 1.2.
Meta released Muse Glimmer on August 10, 2026: 30B parameters, Apache 2.0, 128K context, runs under 20GB at 4-bit. Specs, honest benchmarks and how it compares to Qwen 3.6 and Gemma 4.
Five days after launching a closed paid coding agent, Meta shipped a 30B Apache-2.0 model and announced it will open-source Muse Spark 1.2. What changed, and what it means if you build on these models.
Meta's Muse Code launched August 5, 2026. A head-to-head against Claude Code on benchmarks, pricing, parallel agents, sandboxing and ecosystem — including why Meta's own numbers favour Anthropic.
Meta launched Muse Code on August 5, 2026. What it is, how to install it, what Muse Spark 1.2 costs, how the sandbox and subagents work, and whether the benchmarks hold up.
A verified setup guide for Meta's Muse Code: install, device-code auth, AGENTS.md, MCP, sandbox flags, CI runs and the errors people hit in the first 48 hours.
Meta's contributor tier is 12-21x cheaper because it trains on your code — and it's selected by a single config string. What engineering leaders should do about it.
Meta published a benchmark chart where its own model loses to Claude Opus 5 — and independent harnesses rank Muse Spark 1.2 lower still. A close read of the launch numbers.
Kimi K3 is Moonshot AI's 2.8-trillion-parameter open-weight model. Pricing (API, OpenRouter, free kimi.com tier), specs, license, and Opus 5 comparison.
How Kimi K3 compares to Claude Opus 5, Fable 5, and GPT-5.6 Sol across the Intelligence Index, coding arenas, agentic tasks, and price.
Fable 5's included-in-subscription window ends July 7, 2026. Here's exactly how usage credits work, what Fable 5 costs after the switch, and how to set spend limits so an agent doesn't drain your balance.
Baidu's Unlimited-OCR parses entire multi-page PDFs in a single forward pass. Here's how to run the 3.3B open-weights model locally with Transformers, vLLM, or SGLang.
DSpark is DeepSeek's open-source speculative-decoding module that makes V4-Pro and V4-Flash 51–400% faster — and it works on Qwen3 and Gemma 4 too. Here's how it works and how to use it.
DiffusionGemma 26B-A4B is Google’s first open-weight text-diffusion LLM — a 25.2B MoE built on Gemma 4 that generates text in parallel for up to 4x faster output.
Cohere North Mini Code 1.0 is an open-weight 30B MoE coding model (3B active, 256K context, Apache 2.0) built for agentic software engineering. Specs, benchmarks, access.
A practical GPT-5.6 vs GPT-5.5 comparison: what actually changed across the new Sol/Terra/Luna tiers, pricing, reasoning modes, benchmarks, and a clear decision guide on whether to upgrade or stay put.
OpenAI's GPT-5.6 family — Sol, Terra, Luna — tiers, current pricing after the July 30 cuts, Ultrafast and Cyber, benchmarks, and how to choose.
Claude Fable 5 is back online as of July 1, 2026, after the U.S. lifted its export-control order. Here's what changed, how Anthropic brought it back, and how to access it.