Artificial Analysis vs SemiAnalysis

New Articles: Benchmarking GPT-6 Astra, Intelligence Index v4.3, and more

The articles page now shows 105 articles (up from 94), with new entries including 'Benchmarking GPT-6 Astra' (Sep 9, 2026), 'Announcing the Artificial Analysis Intelligence Index v4.3' (Sep 7, 2026), 'OpenBMB releases MiniCPM5-2B' (Sep 7, 2026), 'Announcing Artificial Analysis Intelligence Index v4.2' (Sep 4, 2026), 'Muse Spark 1.3: Meta reaches the frontier' (Sep 2, 2026), 'Google has released Gemini 3.8 Flash' (Sep 2, 2026), 'Claude Fable 5.1 tops the Artificial Analysis Intelligence Index' (Sep 1, 2026), 'Agnes AI releases Agnes 2.5 Pro Beta' (Aug 27, 2026), 'Intelligence at pocket scale' (Aug 24, 2026), 'Announcing the Speech Agent Arena' (Aug 24, 2026), and 'Announcing the Artificial Analysis Search Index' (Aug 18, 2026).

DeepSeek Launches V4.1 Flash with Novel Encoder-Decoder Architecture

DeepSeek released DeepSeek-V4.1-Flash, a 763B-parameter mixture-of-experts model employing a novel causal encoder-decoder architecture with 8B active parameters for prefill and 16B for decode. It features native vision understanding, 1M token context, an MIT license, and extreme inference efficiency, claiming up to 1/8 KV cache footprint versus V4 Flash. Independent evals (Artificial Analysis Index 40, Vals Index #1 open-weight) show it surpasses V4 Pro at lower cost. API pricing is $0.30/1M input and $1.20/1M output tokens. DeepSeek has soft-retired V4 Pro, routing traffic to V4.1 Flash. The model supports SSD offload and local deployment, with Ollama and Baseten offering day-0 support. Technical discussions highlight the architecture's novelty and potential impact on long-context agents.

• DeepSeek launched V4.1-Flash with a causal encoder-decoder architecture, 763B total params (8B prefill/16B decode active).
• Artificial Analysis Index scores V4.1-Flash at 40, above V4 Pro and below GLM-5.3-Flash.
• API pricing: $0.30 per 1M input tokens, $1.20 per 1M output tokens, cached input $0.006 per 1M.

ModelBest Releases MiniCPM5-2B Edge AI Model

Chinese AI startup ModelBest, in collaboration with the OpenBMB open-source community, has released MiniCPM5-2B, a 2-billion-parameter open-source language model designed for edge devices. The model natively supports tool calling, deep search, code generation, and multi-step reasoning, enabling general-purpose agentic capabilities on resource-constrained hardware. It ranks #1 on the Intelligence Index among open-source models under 4 billion parameters, according to Artificial Analysis, and scores 20 on the Agentic Index. ModelBest has open-sourced the full-stack technical suite, including datasets, training recipes, and reinforcement learning infrastructure, to foster reproducibility. The model aims to shift advanced AI from centralized clouds to edge devices, enhancing privacy, reducing latency, and cutting cloud API costs. Downloads across the MiniCPM family have surpassed 50 million.

• ModelBest released MiniCPM5-2B, a 2-billion-parameter open-source language model for edge devices.
• MiniCPM5-2B ranks #1 on the Intelligence Index among open-source models under 4 billion parameters.
• The model scores 20 on the Agentic Index for autonomous task execution.

SemiAnalysis Tests Engram Offloading to DRAM and SSD

SemiAnalysis analyzes the Engram architecture, a model design that extends standard token embeddings with learned multi-token lookups, allowing for efficient parameter offloading to DRAM or SSD. This reduces HBM capacity requirements for models like DeepSeek-V4.1-Flash. Their experiments show that offloading Engram tables to DRAM can improve performance per dollar, while SSD offloading is currently not economically viable due to overhead. The article also benchmarks inference performance across NVIDIA and AMD GPUs, highlighting NVIDIA's CUDA moat and AMD's slower software support. The analysis includes findings on model behavior, such as gate scan results and ablation studies, and discusses the implications for HBM demand and model architecture innovation.

• SemiAnalysis tested offloading Engram tables to DRAM and SSD for DeepSeek-V4.1-Flash.
• Offloading to DRAM improved performance per dollar, reducing needed HBM capacity.
• SSD offloading was not economically viable, with DRAM delivering 121 million tokens per dollar versus 52 million for SSD.

SemiAnalysis Maps 300 US Datacenter Moratoriums, Finds Minimal Impact

SemiAnalysis published a detailed analysis of datacenter moratoriums in the US, arguing that the widespread narrative that these restrictions are killing the datacenter buildout is inaccurate. The analysis covers over 300 local moratoriums and four state-level actions (New York, Texas, Pennsylvania, Oregon). Using a project-by-project, parcel-level analysis of over 6,000 datacenters, the firm found that only approximately 1,525 MW of planned capacity is genuinely delayed by local moratoriums, representing 7.6% of the capacity sitting inside restricted boundaries. The firm's Datacenter Industry Model forecasts +38 GW of new US datacenter IT capacity in 2027, more than double 2026. The report also discusses public sentiment, finding that 46% of Americans view datacenters unfavorably, and examines the political dynamics driving moratoriums.

• SemiAnalysis maps over 300 local datacenter moratoriums and bans across the US, plus four state-level actions (New York, Texas, Pennsylvania, Oregon).
• Only 1,525 MW of planned capacity is directly delayed by local moratoriums, which is 7.6% of the ~20 GW exposed.
• SemiAnalysis forecasts +38 GW of US datacenter IT capacity delivered in 2027, more than double 2026.

Rubin NVL72 Agentic Inference: 67x Better Performance per Dollar

SemiAnalysis reports first verified agentic inference results for NVIDIA's Rubin NVL72 platform using their AgentX benchmark. Even on early pre-release software, Rubin delivers up to 67x better performance per dollar of TCO compared to GB300 in certain configurations, and significantly higher throughput per MW. The analysis projects Rubin can generate over 2x more profit per gigawatt than Blackwell, with revenue and profit advantages of 39% and 42% respectively at a fixed power budget. Dynamic power shifting (DSX MaxLPS) allows more GPUs per datacenter footprint. The article highlights Rubin's superiority over H200 and MI355X, with recommendations for inference providers to adopt Rubin for cost-efficient token generation.

• Rubin NVL72 achieves up to 67x the throughput per TCO of GB300 in specific scenarios.
• Rubin delivers up to 7x better token throughput per MW than Blackwell in real-world tests.
• Rubin can generate over 2x more profit per gigawatt than Blackwell.