DeepSeek-V4 Arrives with 1M-Token Context Support — Hybrid Attention Delivers Major Inference Efficiency Gains
DeepSeek has released its V4 series. It includes two Mixture-of-Experts (MoE) models — **DeepSeek-V4-Pro** (1.6T parameters, 49B active) and **DeepSeek-V4-Flash** (284B parameters, 13B active) — and both support a **context length of 1 million tokens**.
The biggest technical highlight is the **Hybrid Attention Architecture**. By combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), it reportedly reduces per-token inference FLOPs to 27% and KV cache to 10% compared with DeepSeek-V3.2 at the 1M-token setting.
The release also introduces **Manifold-Constrained Hyper-Connections (mHC)**, which strengthen conventional residual connections, and the **Muon Optimizer** for improved training stability. After pretraining on more than 32T tokens, the models adopt a two-stage post-training pipeline that first develops domain-specific experts independently and then integrates them through on-policy distillation.
On benchmarks, DeepSeek-V4-Pro-Max delivers top-tier performance among open-source models, ranking at the top of coding benchmarks and significantly closing the gap with closed-source frontier models on reasoning and agent tasks. The Flash variant can reportedly match the Pro version's reasoning performance in some cases when given a larger thinking budget.
On Hugging Face, the Pro version has already earned 4.7k likes and the Flash version 1.4k — a clear sign of how closely the community is watching.
NVIDIA and the UK Expand Their Sovereign AI Foundation — Heavy Infrastructure Investment Centered on Isambard-AI
At London Tech Week, NVIDIA announced that the UK's sovereign AI initiative is starting to deliver concrete results. A year after Jensen Huang and Prime Minister Keir Starmer declared that the UK would become an "AI maker," the number of AI cloud provider deployments in the UK has doubled.
Nebius announced three new NVIDIA AI infrastructure deployments, with plans to reach 65 megawatts by 2027. BT and Nscale will build sovereign AI data centers at three BT sites across the UK. Meanwhile, **Isambard-AI** — the UK's largest computer, built from 5,400 NVIDIA GH200 Grace Hopper Superchips — is running on zero-carbon power and underpinning the nation's AI research.
Notably, investment in sovereign AI goes beyond infrastructure buildout, with applications advancing across a broad range of domains including healthcare and drug discovery, coding, and agentic AI.
SIA: Self-Improving AI Enters New Territory by Combining Harness and Weight Updates
**SIA (Self Improving AI)**, introduced on Hugging Face, proposes a self-improvement loop that updates an agent's harness/scaffold and the model's weights at the same time. Most previous self-improvement approaches either changed only the scaffold or only the weights; SIA runs both.
What makes it interesting is that the two levers produce very different effects. Harness updates mainly improve the "software engineering around the model" — parsing, retries, tool usage, retrieval procedures — while weight updates tend to add task-specific intuition that could not be discovered through the scaffold alone.
In experiments, it outperformed a harness-only approach on three tasks: legal classification, CUDA kernel optimization, and scRNA denoising. It is drawing attention as a promising research direction for self-improving AI.
A New Approach to Making LLM Reasoning Transparent with Causal Graphs
Another study drawing attention on Hugging Face presents a method for **LLM explainability using counterfactual chains and causal graphs**. It flips the usual idea: instead of using LLMs to extract causal graphs of the external world, it uses causal graphs to model the LLM's own reasoning process.
Concretely, it discovers human-interpretable concepts and maps inputs to concept states, generates counterfactual chains with an MCMC-style extension, and constructs a causal graph with σ-CG. When three LLMs were evaluated on three tasks — disease diagnosis, sentiment analysis, and LLM-as-a-judge — the learned graphs showed high predictive fidelity and structural stability.
As explainability work on the black-box nature of LLMs grows in practical importance, this approach could become a practical diagnostic tool.
MMAE: A Comprehensive Benchmark for Audio Editing Exposes the Limits of Current Models
**MMAE (Massive Multitask Audio Editing)** is the first comprehensive evaluation benchmark for instruction-based audio editing. It spans 7 audio modalities — including sound, speech, music, and their mixtures — 6 levels of task complexity, and 8 operation types.
It comes with 2,000 high-quality samples curated through human-agent collaboration and a rubric-based evaluation framework (17,741 verifiable criteria). Evaluations of state-of-the-art models produced a striking result: **Exact Match Rate (EMR) was consistently below 5%**, dropping to 0% on complex mixed-modality tasks — laying bare how far current audio editing models remain from reliable editing.
FT: Software Acquisitions Hit Their Lowest Level Since the Pandemic — The Impact of the AI Bubble Correction
According to the Financial Times, M&A deals for software companies have sunk to their lowest level since the pandemic. The so-called "AI route" (AI-driven market disruption) is in the background, and the effect of investors rethinking software valuations appears to be significant.
This is being watched as a sign that the exuberance of the AI boom is entering a correction phase. Whether it is a short-term deterioration in market sentiment or the beginning of a structural reevaluation of AI investment, future trends will be the focus.
FT Opinion: AI Agents Should Not Be Granted Legal Personhood
The FT also ran an opinion piece opposing the granting of legal personhood to AI agents. As AI increasingly acts autonomously, debate over clarifying where legal responsibility lies is heating up.
Granting legal personhood may at first glance make transactions more efficient, but it carries the risk of blurring ultimate responsibility, shifting it ambiguously from humans to AI. How to design the balance between AI autonomy and legal accountability is likely to become a key regulatory issue.
Qiita: Running SurrealDB as the "Organizational Memory" for AI Agents
A Qiita post introduced an operational case study of using SurrealDB as a memory and collective-intelligence foundation for a fleet of AI agents. By preserving "previous judgments" — which tend to be lost when working with LLMs — in SurrealDB's structured data and full-text search, the setup has already accumulated more than 7,000 records.
It also incorporates a monthly review process, making it a useful source of practical insight on how to design and manage long-term memory for AI agents.
Qiita: The "Quick PoC Disease" That AI Development Easily Falls Into
Another Qiita post introduced the concept of "quick PoC disease" in AI development. PoCs (proofs of concept) matter in themselves, but the term refers to the state of mass-producing PoCs without ever transitioning to production. The point that the faster AI validation becomes, the easier it is to fall into this trap, is a painful one for practitioners.