TiTikey

DeepSeek released a set of new open-source models from the long-awaited V4 family that perform on par with GPT-5.4 and Opus 4.6. - DeepSeek-V4 Preview - DeepSeek-V4-Pro - DeepSeek-V4-Flash V4 models are available on APIs and DeepSeek Chat! Testing time 👀

DeepSeek-V4 just dropped! And it's solving one AI's biggest problem today: It runs 1M-token context at 10% of the KV cache and 27% of the inference FLOPs of V3.2. Here's what that means. KV cache is the memory footprint your GPU holds for every token already in context. It grows linearly with context length, and at 1M tokens it's usually what forces you onto bigger hardware or kills your throughput. Cutting it to 10% means you can serve longer contexts on smaller machines, or pack far more concurrent users on the same ones. Inference FLOPs is the compute cost of generating the next token. With vanilla attention this scales quadratically with context length, which is why long contexts get brutally expensive per token. 27% means each generated token at 1M context is nearly 4x cheaper to produce. Put together, long-context inference goes from a premium feature you ration to something you can run by default. The trick is a hybrid attention design that interleaves two mechanisms instead of picking one. 𝗖𝗦𝗔 (𝗖𝗼𝗺𝗽𝗿𝗲𝘀𝘀𝗲𝗱 𝗦𝗽𝗮𝗿𝘀𝗲 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻) first squashes every 4 KV entries into a single compressed entry. Then it uses a lightning indexer to select the top-k most relevant compressed blocks. Compression and sparsity stacked. 𝗛𝗖𝗔 (𝗛𝗲𝗮𝘃𝗶𝗹𝘆 𝗖𝗼𝗺𝗽𝗿𝗲𝘀𝘀𝗲𝗱 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻) goes aggressive. It compresses every 128 tokens into one entry and skips sparse selection entirely, because at that compression ratio dense attention over a tiny set is already cheap. Both get a sliding window branch over the last 128 tokens, so local fine-grained structure isn't lost to compression. The result is that CSA handles medium-grained retrieval while HCA handles coarse context summarization, and alternating them across layers gives you both without paying for both at full cost. V4-Pro (1.6T total, 49B active) ranks 23rd among human Codeforces competitors and hits 120/120 on Putnam-2025. Open weights on Hugging Face. The era of million-token context in open models has effectively started.

Babe Wake Up! DeepSeek just mass-murdered OpenAI's pricing model with genius-level AI at 7x cheaper. The benchmarks shows DeepSeek-V4-Pro-Max crushes GPT-5.4-xHigh and Claude-Opus-4.6-Max in reasoning and coding. But here's the real story: economics. OpenAI and Anthropic charge $15-$25 per million output tokens. DeepSeek delivers comparable-often superior-intelligence for $3.48. The comparison: - GPT-5.5: $5.00 in / $30.00 out - GPT-5.4: $2.50 in / $15.00 out - Claude Opus 4.7: $5.00 in / $25.00 out - DeepSeek: $0.87 in / $3.48 out 1M context window. Native thinking modes. Completely open source. DeepSeek is back on the game.

The pricing floor for Opus-tier and GPT-tier reasoning just dropped, if you're willing to self-host. DeepSeek V4 Pro activates 49B parameters out of 1.6T per token. It posts Codeforces 3206 (legendary grandmaster tier), beats Opus 4.6 on Apex Shortlist (90.2) and SimpleQA (57.9 vs 46.2), and ties on SWE Verified. Open weights. 1M context. Sparsity ratio of 32 to 1. Closed frontier is running dense or near-dense with likely 5-10x the active params. Inference cost tracks active params almost linearly. Any SaaS paying $15-60 per million tokens for Opus or GPT-5.4 now has a credible self-host option at a fraction of that cost. Where closed labs still clearly lead: Terminal Bench (GPT +7), Toolathlon (GPT +3), HLE (Gemini +7), SimpleQA (Gemini +18). Agentic tool use and deep knowledge synthesis. Those require RL infrastructure and proprietary data pipelines that don't commoditize the way raw math and coding do. The commodity layer expanded. The moat layer narrowed to two things.

Andrej Karpathy could have charged $10,000 for this course. He put it on YouTube. The man who built Tesla Autopilot from scratch. Co-founded OpenAI. Understands AI at a level most engineers at Google and Meta never reach. Sat down. Recorded 2 hours. No frameworks. No libraries. No shortcuts. Then dropped it for free. The gap between people who watch it this week and those who save it for later is not 2 hours. It is everything those 2 hours quietly unlock for the rest of your career.

deepseek v4 is now the cheapest sota model available at 1/20th the cost of opus 4.7. for perspective, if uber used deepseek instead of claude their 2026 ai budget would have lasted 7 years instead of only 4 months.

I asked Opus 4.7 and GPT-5.5 about DeepSeek's latest launch. Opus: The AI we're trusting for research doesn't even know what JUST dropped

DeepSeek releases DeepSeek-V4. 🐋 - DeepSeek-V4-Pro: 1.6T params - DeepSeek-V4-Flash: 284B params DeepSeek-V4-Pro rivals Claude-Opus-4.6-Max, GPT-5.4-xHigh and Gemini-3.1-Pro-High. They support 1M context length, thinking and set new records for Codeforces.

Summary of GPT-5.5: - Terminal Bench 2.0: 82.7% accuracy - Stronger than Opus 4.7 + Gemini 3.1 - Builds full video games - Creates full landing pages from prompts - Controls apps using computer-use mode - Runs long-horizon coding tasks Every day there’s improvements.

Nvidia CEO Jensen Huang: "I'd hire the graduate who's expert in AI over the one who isn't. Every time"

Big drop for GBrain v0.19. Skillify is now at full strength as a skill that you can use with GBrain! If you're already using GBrain with OpenClaw/Hermes just say "Upgrade GBrain" to take advantage of it.

Big drop for GBrain v0.19. Skillify is now at full strength as a skill that you can use with GBrain! If you're already using GBrain with OpenClaw/Hermes just say "Upgrade GBrain" to take advantage of it.

An entire robot vibe coded with GPT-5.5 in minutes… This used to take weeks to make The builders are going to cook with 5.5