NEW DeepSeek V4 Flash Update!
DeepSeek V4 Flash has been officially released with significant performance improvements over its preview version, designed specifically for AI agents. The speaker demonstrates practical applications by building 50+ projects and integrating it into their agent operating system, while emphasizing it's not frontier-level but offers excellent speed and cost efficiency.
Summary
DeepSeek announced a major upgrade to V4 Flash, transitioning from preview to official API status. The benchmark improvements are substantial: AIME increased from 61.8 to 82.7, NL2 repo from 39.4 to 54.2, Cyber Gym from 38.7 to 76.7, and Deep SWE from 7.3 to 54.4. The model performs comparably to Claude Opus 4.8 but is positioned below frontier models like Fable 5. Critically, DeepSeek achieved these improvements through better training rather than increasing model size—V4 Flash maintains the same architecture and size as its preview version. The speaker tested the model extensively by building over 50 projects in a single day, including interactive games and 3D environments, demonstrating its capability for complex logic and UI design. The model is specifically optimized for agentic work, making it ideal for AI agent frameworks like Hermes. Integration into the speaker's agent operating system took minutes, allowing the model to slot into existing workflows alongside Claude and other models. V4 Flash operates with a 1 million token context window and is 12% more token-efficient than its predecessor. Pricing is significantly lower than competitors—60% cheaper than GPT-5.6 and Luna Max. The speaker clarifies that the upgrade is API-only; the deepseek.com web interface and V4 Pro remain unchanged. Open weights are expected to release soon, which would make V4 Flash the second-highest scoring open-weight model behind Llama 3. The speaker emphasizes building flexible AI systems where new models can be plugged in immediately upon release rather than chasing individual models.
Key Insights
- DeepSeek achieved significant performance upgrades to V4 Flash through improved training methodology rather than increasing model size or parameters, maintaining identical architecture to the preview version
- V4 Flash is specifically tuned for agentic work and agent frameworks like Hermes, not designed as a flagship coding model, representing a different optimization priority than frontier models
- The speaker built over 50 functional projects including 3D games in a few hours with V4 Flash, demonstrating practical capability for complex logic, UI design, and rendering despite not reaching frontier model quality
- V4 Flash is priced 60% lower than GPT-5.6 and comparable frontier models while achieving performance within one point of GLM 512 and comparable to Gemini 3.6 Flash
- The upgrade is API-only as of the release date; the deepseek.com web interface and V4 Pro API remain unchanged, with V4 Pro expected to release separately as a potential frontier-level model
Topics
Transcript
[0:00] A brand new version of DeepSeek, DeepSeek 4 Flash just went live and is built for AI agents. You can see the announcement just happened a few hours ago today. DeepSeek just dropped a major upgrade to V4 Flash and makes your agents seriously more powerful. So, DeepSeek say the new benchmark scores far surpass their previous top preview model from the small, fast, and cheap tier. I'll show you exactly what I've built with We actually built over 50 things with it already today. So, we've tested it relentlessly and that means your agents get to think across a [0:30] million tokens of context, they run longer coding loops, and finish more work uh before you touch your…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Julian Goldie SEO
Google AI Studio Just Got A Huge Security Upgrade
Google is reportedly developing a new "Security Check" mode for AI Studio that would apply secure coding practices to AI-generated applications. The feature aims to automate security verification during app development, addressing common vulnerabilities like exposed keys, weak authentication, and data access issues before deployment.
New Gemini Skills Update Is WILD!
Google has launched Gemini Skills, a new feature allowing users to save custom instructions once and reuse them repeatedly by typing a slash command, replacing the discontinued Gems feature. Skills can include reference files, be combined together, and are built on an open file format (skill.md) that allows portability across different AI tools.
NEW Codex CLI Update Is ABSURD!
OpenAI's new Codex CLI update transforms terminal-based programming assistance by enabling multiple AI agents to work simultaneously on different tasks, visible in a full-screen control center interface. The update includes features like Agent View, work trees, voice control, and parallel task execution, fundamentally changing how AI assistance operates from sequential Q&A to autonomous team-based work management.
NEW Hermes Cloud + Agent OS is Crazy! 🤯
Hermes Cloud combined with Agent OS enables users to deploy AI agents in the cloud 24/7 using free models without API keys or local setup complexity. The system provides persistent memory, sandbox isolation, and integration with custom workflows like content research, competitor analysis, and news monitoring.
Manus 2.0 Just Changed AI Agents Forever
Manas 2.0 represents a fundamental architectural overhaul that enables users to create games, websites, videos, and automations through simple text requests, with the key innovation being editable components rather than fixed outputs. The platform uses a new Cascade engine that reduces token usage by 23% and improves task speed by 28%, while maintaining all project elements in one connected workspace.