DeepSeek V4-Flash: Small Model, Big Pricing Pressure
DeepSeek’s compact V4-Flash undercuts larger models on cost—builders take note. Plus: EU AI labels now mandatory, and a 35B LLM on €870 hardware.
The story of the day
DeepSeek’s new V4-Flash model is the latest proof that the cost curve for capable LLMs is bending. The smaller variant of its flagship V4, released Friday, matches larger open-source and proprietary models in early tests while undercutting them on price. For developers, this isn’t just another benchmark chase—it’s a pricing signal. Closed-source incumbents have leaned on performance-per-dollar as a moat; open-weight models like V4-Flash erode that by offering parity at a fraction of the cost. The question for builders: how long until inference budgets stop being a line item and start being a rounding error.
New tools & launches
Galda — Git-agnostic Claude Code/Codex project switcher. Open source. [Source]
TamedTable — Natural language ETL harness for LLMs. Open source. [Source]
Abloatai — Multiplayer AI infrastructure framework. Open source. [Source]
Models & research
DeepSeek V4-Flash (open-weight) — A distilled variant of DeepSeek’s V4, early tests show it punches above its weight class, delivering performance comparable to larger models at lower cost. The release underscores a trend: open-weight models are closing the gap on proprietary offerings by optimizing for efficiency, not just scale. [The Information AI]
35B LLM on €870 Hardware — A proof-of-concept demo runs a 35B-parameter model at 128K context on used consumer GPUs, no cloud required. The setup uses quantization and offloading tricks, but the takeaway is clear: the barrier to running large models locally is dropping fast. [Hacker News AI]
Industry pulse
The EU’s new rule mandating AI labels on authentic-looking content takes effect. Publishers and platform builders now face compliance costs, but the bigger friction may be user trust—labels are only useful if audiences understand what they mean. [The Guardian]
Stat of the day
€870 — The cost of used hardware sufficient to run a 35B-parameter LLM at full speed and 128K context, per a recent demo. For context, that’s less than a single high-end GPU’s retail price. [Hacker News AI]
Get the full analysis — subscribe to Gradient Pro.
