DeepSeek V4.1 Flash: 552B MoE, 1M Context, $0.30 Input, MIT License
DeepSeek V4.1 Flash is a 552B MoE with 1M context, MIT weights and $0.30 per million input tokens. Benchmarks, third-party scores and self-hosting math.
·15 min read
Topic
LLM
9 articles
AIGeopoliticsTech IndustryDeveloper ToolsCybersecurityInfrastructureAI InfrastructureAI ModelsSemiconductorsCloud InfrastructureWeb DevelopmentBreaking News
Kimi K3: China's 2.8T Open-Weight Model Beats GPT-5.6 Sol in Code
Moonshot AI released Kimi K3 on July 16 — a 2.8T sparse MoE model ranking first globally in Frontend Code Arena at 1,679 points. Full weights drop July 27.
·10 min read
Grok 4.5: SpaceX Paid $60B for Cursor and Built This Model
xAI launched Grok 4.5 on July 8 at $2/$6 per million tokens — 60% cheaper than Opus 4.8. SpaceX acquired Cursor for $60B in June to build it. Benchmarks and developer guide inside.
·9 min read
OpenAI GPT-5.6: Sol, Terra, and Luna Pricing, Benchmarks July 2026
OpenAI launched GPT-5.6 Sol, Terra, and Luna June 26 to 20 vetted partners. Sol is $5/$30, Terra $2.50/$15, Luna $1/$6 per 1M tokens. General access expected mid-July.
·9 min read
Gemini 3.5 Pro: 2M Context, Deep Think, and When It Launches
Gemini 3.5 Pro hits GA in late June 2026 with a 2M token context and Deep Think mode. Specs, pricing, and comparison to Fable 5 and GPT-5 explained.
·10 min read
Microsoft Launches 7 In-House AI Models at Build 2026 to Cut OpenAI Reliance
Microsoft unveiled 7 in-house MAI models at Build 2026: MAI-Thinking-1 (35B active params, no OpenAI data) and MAI-Code-1-Flash now live in GitHub Copilot and VS Code.
·8 min read
Gemini 3 Deep Think Hits 84.6% on ARC-AGI-2 and 3455 Elo on Codeforces
Google's Gemini 3 Deep Think scored 84.6% on ARC-AGI-2, 48.4% on Humanity's Last Exam, and 3455 Elo on Codeforces. Gemini 3.1 Pro is now in preview. Here is what the benchmarks actually mean.
·8 min read
Xiaomi's Hunter Alpha: The 1-Trillion-Parameter Model That Ran Anonymous for 8 Days
On March 11 a mystery 1-trillion-parameter model appeared on OpenRouter. The AI community burned 500 billion tokens assuming it was DeepSeek V4. On March 19 Xiaomi revealed it was theirs.
·7 min read
RAG in Production 2026: Chunking Strategies, Embedding Costs, and What Actually Works at Scale
Most RAG tutorials show you how to build a demo. This post covers what breaks in production: chunking at 512 tokens beats semantic splitting, embedding costs range from $0.02 to $0.18 per million tokens, re-ranking boosts precision by 18–42%, and agentic RAG is now the 2026 standard. A practical guide for developers shipping RAG to real users.
·12 min read