Stepfun Step 5 Preview (LLM) lands on Artificial Analysis Pareto frontier

Social·Community

StepFun joins the frontier: a previously non-frontier Chinese lab (StepFun) released a Kimi K3-level model, 3 times cheaper per Artificial Analysis

Chinese lab StepFun released Step 5 Preview, which a Reddit post says scores 44 on Artificial Analysis' index, equal to Kimi K3 (max), at one-third the price. The model sits on the cost/intelligence Pareto frontier, marking StepFun's entry into the frontier tier. The poster notes it is closed-source, leaving its competitiveness against other Chinese labs uncertain.

Chinese lab StepFun released Step 5 Preview, which a Reddit post says scores 44 on Artificial Analysis' index, equal to Kimi K3 (max), at one-third the price. The model sits on the cost/intelligence Pareto frontier, marking StepFun's entry into the frontier tier. The poster notes it is closed-source, leaving its competitiveness against other Chinese labs uncertain.

Developments·Benchmark

Artificial Analysis updates its Gemini 4 Argon model benchmark page

Artificial Analysis updates its benchmark for Gemini 4 Argon

Artificial Analysis updates its benchmark for Gemini 4 Argon

Social·Community

K2 Horizon lineup lands on Artificial Analysis; KV cache design inflates RAM needs despite SOTA Intelligence Index scores

A Reddit analysis of the newly listed K2 Horizon model lineup on Artificial Analysis argues the AA intelligence-vs-parameters plots mislead, because the models' KV cache design inflates RAM needs at long context. The poster maps each tier (3.7B, 7B, 36B-A4B) to the specific hardware where it makes sense, often concluding alternatives like Qwen3.8-27B or MiniCPM5-2B are better buys.

A Reddit analysis of the newly listed K2 Horizon model lineup on Artificial Analysis argues the AA intelligence-vs-parameters plots mislead, because the models' KV cache design inflates RAM needs at long context. The poster maps each tier (3.7B, 7B, 36B-A4B) to the specific hardware where it makes sense, often concluding alternatives like Qwen3.8-27B or MiniCPM5-2B are better buys.

Releases·Model Releasefeatured

DeepSeek V4.1 Flash is getting surprisingly close to GPT-5.6 Sol territory, while being absurdly cheap

DeepSeek released V4.1-Flash, an open-weight (MIT) MoE model built on a new causal encoder-decoder architecture with asymmetric active parameters — 8B for input/prefill and 16B for output/decode. The release adds vision input, 1M-token context, and aggressive KV-cache compression, drawing independent praise as the new top open-weight model on cost-adjusted intelligence. It succeeds V4 Pro, which DeepSeek is quietly retiring, and several commentators argue the '0.1 version' label understates a fundamental architectural overhaul.

DeepSeek released V4.1-Flash, an open-weight (MIT) MoE model built on a new causal encoder-decoder architecture with asymmetric active parameters — 8B for input/prefill and 16B for output/decode. The release adds vision input, 1M-token context, and aggressive KV-cache compression, drawing independent praise as the new top open-weight model on cost-adjusted intelligence. It succeeds V4 Pro, which DeepSeek is quietly retiring, and several commentators argue the '0.1 version' label understates a fundamental architectural overhaul.

Social·Community

DeepSeek 4.1 Flash uses 88,574 output tokens per task vs GPT-6 Astra Low's 4,432 at lower intelligence index

A Reddit user highlighted Artificial Analysis data showing DeepSeek 4.1 Flash consumes far more output tokens per task than comparable models. At 88,574 output tokens for a 40 intelligence index score, it uses roughly 20x the tokens of GPT-6 Astra Low, suggesting the model generates much more internal reasoning to reach similar results. The caveat: output-token counts are a proxy, not a direct quality or cost measure.

A Reddit user highlighted Artificial Analysis data showing DeepSeek 4.1 Flash consumes far more output tokens per task than comparable models. At 88,574 output tokens for a 40 intelligence index score, it uses roughly 20x the tokens of GPT-6 Astra Low, suggesting the model generates much more internal reasoning to reach similar results. The caveat: output-token counts are a proxy, not a direct quality or cost measure.