Full DeepSeek V4 Paper Reveals Complete Details of FP4 Quantization-Aware Training

DeepSeek released the full paper this week, a substantial expansion of the earlier V4 preview (58 pages). The biggest highlight is the implementation details of **FP4 quantization-aware training (QAT)**.

How FP4 QAT Works

The expert weights in the MoE (Mixture of Experts) are quantized directly to FP4, with QAT carried out in the later stages of training. Rather than the conventional “quantize after training” approach, this method optimizes the training itself with quantization in mind. The CSA indexer in the QK path also adopts FP4 activations, achieving a 2x speedup while maintaining 99.7% recall. Inference runs directly on FP4 weights.

Dramatic Efficiency Gains

The efficiency comparison table in the paper is striking. With V3.2 as the baseline, V4-Pro cuts FLOPs to 27% and KV cache to 10%. V4-Flash goes even further, compressing FLOPs to 10% and KV cache to just 7%.

Training Stability Techniques

The paper documents two mechanisms for tackling the “loss spike problem” that frequently plagues trillion-parameter-scale MoE models.

**Anticipatory Routing** — Deliberately desynchronizes updates between the main model and the router, preventing feedback loops from amplifying anomalies. It adds 20% overhead but only kicks in during loss spikes

**SwiGLU Clamping** — Applies hard limits to SwiGLU's linear path (-10 to 10) and gate path (max 10) to suppress cascades of extreme values

Generative Reward Model

Instead of a separate reward model for RLHF, DeepSeek uses an approach where the generative model itself performs the evaluation. Trained on scored data, it achieves self-evaluation with reasoning. This keeps human labeling to a minimum while establishing a unified training pipeline.

Evaluation Results

In Chinese text generation, V4-Pro posted a 62.7% win rate against Gemini 3.1 Pro (77.5% on writing quality). In coding agent evaluations, 52% of users said V4-Pro could be adopted as their default coding model. Reddit discussions include multiple comments along the lines of “I switched to V4-Pro last week and haven't noticed any quality drop in my day-to-day work.”

The fact that FP4 QAT keeps quality degradation to a minimum could translate directly into cost savings, especially in multi-agent configurations.

---

Emotion AI Is “Quietly Seeping” Into the Workplace — The Atlantic Investigation

An investigation by The Atlantic's Ellen Cushing lays bare how software that **claims to read human emotions with AI** is quietly but steadily spreading through the workplace.

Notably, The Decoder's coverage of the article uses the word “pseudoscientific.” The scientific basis of emotion-recognition AI has been debated for years, and prominent AI researchers have voiced skepticism. Yet actual deployment is racing ahead of the theoretical debate.

The issue exemplifies the structural problem of AI's societal impacts being “deployed before they are regulated.” Surveillance-style AI use in the workplace is facing tighter restrictions under the EU AI Act and similar regulations, but how already-deployed systems actually operate remains difficult to see from the outside.

---

Big Tech's AI Investment Reaches $725 Billion as Free Cash Flow Sinks to Decade Lows

According to the Financial Times, AI-related investment by major tech companies has hit **$725 billion**, and as a result, free cash flow has fallen to its lowest level in the past decade.

The sheer scale of the investment stands out. This is not mere infrastructure spending — it spans every front of the AI race, including GPU purchases, data center construction, and talent acquisition. Meanwhile, when and at what scale returns on this investment will materialize remains opaque, and wariness is growing in the markets.

This move is also the flip side of Cloudflare's recent cut of 1,100 jobs (attributed to AI-driven efficiency). Those making massive investments on one side, and those reducing headcount through efficiency on the other — AI's economic impact is rippling through the industry asymmetrically.

---

“Clients Used to Demand Carousels, Now It's AI Chatbots” — The Cycle of AI Fads

A notable essay that racked up 69 points on Hacker News. It paints a satirical picture of the fad cycle in web development: “a few years ago every client demanded carousels, then parallax, and now AI chatbots.”

Its candid, experience-based analysis of how tech trends drive client demands — and the gap with actual user value — has struck a chord. Across its 22 comments, people are debating the fundamental question of “how many sites actually need an AI chatbot?”

As a healthy dose of skepticism toward the air of inevitability surrounding the AI boom, it's a discussion practitioners will find worthwhile.

---

The Limits of LLM-as-judge in Code Evaluation

An article that gained traction on Hacker News points out the problems of using LLMs as “judges” for code evaluation and proposes alternative approaches.

LLM-as-judge is widely used in RLHF and automated evaluation, but when it comes to code, the argument goes, there is a clear ground truth — does it work or not — so test suite results should take priority over LLM judgments. In particular, LLM judgments have been reported to lack consistency when evaluating edge cases and performance.

The piece is a reminder of the importance of understanding the limits of LLM evaluation and designing an appropriate evaluation pipeline on top of that.