Search “Reasoning”

Cerebras on the MCP vs. CLI Debate: Inference Infrastructure and Code Execution Environments Are Equally Critical

Perplexity CTO Denis Yarats announced at Ask 2026 that the company is abandoning MCP and returning to APIs and CLI, sparking the MCP vs. CLI debate. Since Anthropic open-sourced MCP in November 2024, it has surpassed 97 million monthly downloads, but connecting the three services GitHub, Slack, and Sentry can consume over 55,000 tokens before the user's message, with token usage 3-42 times higher than CLI. Cerebras points out that its Wafer-Scale Engine can deliver up to 3,000 tokens/second, roughly 15 times the inference speed of GPU-based solutions, while Pydantic's Rust interpreter Monty starts up in under 0.06ms.

Cerebras: Fast Inference Finds Its Rhythm

Cerebras looks back at 2025: DeepSeek R1 Llama 70B runs at over 1500 tokens/second on its chips, about 57 times that of GPUs; independent testing by Artificial Analysis shows its inference performance exceeds NVIDIA Blackwell's record. Cognition's SWE-1.5 and SWE-grep series achieve frontier-level coding performance at up to 13 times the speed of general-purpose models. Cerebras has launched a 40+ exaflop data center in Oklahoma and is deploying more than 7 data centers across North America and Europe.

Baseten Engineering Blog: Inference Engineering from GPT-2 to Kimi K3 and Open-Source LLM Training

The Baseten Engineering Blog's Foundations section features two articles: one traces the evolution of inference engineering from GPT-2 to Kimi K3, and the other explains why open-source LLM training is chaotic and how it works.

Baseten Engineering Blog: LLM Inference Observability, DeepSeek V4 Pro 0813, and NVIDIA Nemotron 3.5

The Baseten engineering blog lists several inference engineering and model articles, including "How Leading Platforms Ensure LLM Inference Observability," inference engineering for DeepSeek V4 Pro 0813, and Baseten providing open-source inference for You.com. The blog also introduces NVIDIA Nemotron 3.5 Lightning and Nemotron 3.5 ASR Streaming, and features content such as Laguna S 2.1, Qwen3-TTS fine-tuned voice cloning, and Kimi K3 routing.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

The Father of OpenAI Reasoning Noam Brown's Latest Interview: Mathematics Is Just an Appetizer for the Multi-Agent Era

OpenAI researcher and core o1 author Noam Brown said in the latest podcast that multi-agent systems are a way to change inference-time compute from serial scaling to parallel scaling, but in solving the Millennium Prize Problems, multi-agent systems account for at most 10% of the credit; the core is still that the model itself is sufficiently powerful. He revealed that when 5.6 was released, they had a truly formed multi-agent system for the first time, provided in the form of Ultra Mode, with 4 Agents by default and adjustable higher; with 4 Agents, the problem-solving speed is about twice as fast, and tests scaled up to as many as 16 Agents.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.