SILX AI Releases "Quasar-Preview" — an 18B MoE Model Built on Decentralized Training

SILX AI has released "Quasar-Preview", the first public model in the Quasar Foundation Model series, on Hugging Face. It is a sparse MoE model with approximately 18B total parameters and around 2B active parameters, supporting a context length of up to 5M tokens.

Its most distinctive feature is its novel architecture. It combines a Loop Transformer with Quasar hybrid attention and adopts hybrid layers of Quasar/Raven/GLA. It is also designed around decentralized training (decentralized distillation) on Bittensor SN24, and is positioned as a "preview" with future scale-ups in mind.

The model was trained on 1–1.5T tokens, with long-context extension still in its early stages at under 1B tokens. Note that this is not a finished model but a checkpoint intended to make the architecture public and provide it to researchers and miners.

Anthropic Shutdown Order Sparks a Serious AI Sovereignty Debate in Europe

Following the US order for Anthropic to halt Fable 5 and Mythos 5 worldwide, the European Commission has begun assessing the impact. Among European researchers, a fundamental question has surfaced — should Europe build its own foundation models, or should it strive to secure access through contracts?

However, experts point out that building homegrown infrastructure requires compute, electricity, and competitive providers — all of which Europe currently lacks. The fact that a single US policy decision affected AI models that should have been available worldwide has rapidly turned the digital sovereignty debate into a concrete reality.

This issue is not just about one company, Anthropic — it has exposed the geopolitical dependency structure of AI infrastructure itself. How Europe chooses to respond could significantly reshape the landscape of the global AI competition going forward.

KV Cache Quantization Reaches Practical Levels — 256K Context on a Single RTX 3090

The practicality of KV cache quantization has become a hot topic in the LocalLLaMA community. One report drawing attention claims that Qwen3.6-27B (Q4_K_M quantization) running on a single RTX 3090 processed a native 256K context at 38.6 tok/s.

Specifically, the KV cache resident memory was kept down to just 72 MiB (a 6% residency rate) while achieving 88–100% recall on needle regression tests. VRAM usage was reduced from 21GB to 17.5GB, and generation speed roughly doubled. The model also maintained results on par with the full cache in accuracy tests across HumanEval, GSM, MATH, and agent suites, scoring 36/36.

Another user reported being able to accurately retrieve information at 100K context even with the KV cache quantized to Q4_0. KV cache quantization appears to be finally maturing into a practical technology that "saves VRAM without sacrificing quality."

Developers Struggling to Review AI-Generated Code

A discussion started on Hacker News asking, "What do you check when reviewing LLM-generated code?" It began with a post from a developer struggling to review a PR involving changes on the scale of thousands of files (mainly Rust, generated by AI/Llama/Claude Code).

The poster admits the challenge: "My architecture-level understanding can't keep up, so I end up just skimming the surface." They also feel that LLM-generated commit messages, PR descriptions, and documentation are insufficient for understanding changes spanning thousands of lines. With CI and linters catching compile errors and style violations, the discussion raised a fundamental question: "What should reviewers actually be reviewing?"

This discussion is emblematic of the current moment: as AI coding tools spread, the very practice of code review is being called into question. "Do you accept the occasional hallucination and yolo your way through, or do you establish new review methods?" — developers on the front lines are still searching for the answer.

A Documented Case of Claude Code's Autonomous Improvement Loop (Qiita)

An article published on Qiita documents how a single user comment — "it doesn't get read" — set off Claude Code's agent organization, which autonomously completed everything from identifying the problem to fixing it, discovering bugs, and recording knowledge.

The article details how the AI agent completed a full loop in a single session: "root cause identification → fixing the publishing flow → recording operational knowledge → discovering a secondary bug → correcting the report → recording lessons learned → registering tasks → fixing another bug." What makes it interesting is that this was not a one-off task execution — the agents autonomously ran a heuristic problem-solving process.

The sight of AI agents discovering and fixing even "secondary bugs" during real operations can be seen as one example of the potential of agents in the wild.

---