중국 미니맥스와 딥시크, H100급 10만개 미만 GPU로 최전방 모델 확장…디스틸레이션으로 설명 못해

The heretofore uncontested placement of the DeepSeek-V4.1-Flash-1017 model at the pareto on the agentic benchmark score vs. cost curve (which simply means that the model achieves an optimal balance between the two) was fairly short-lived, with the new MiniMax M3.1 Flash having just taken a lead. Even so, with Anthropic and OpenAI both clamping down on distillation attempts, it is uncertain if China's AI labs would be able to maintain their current trajectory, especially given their relatively minuscule compute footprint. The MiniMax M3.1 Flash model debuts with 428 billion total parameters and just 23 billion active ones, and priced at just $0.15/$0.60 per 1 million input/output tokens The MiniMax M3.1 Flash model offers frontier-level capabilities at an extremely low cost, with 1 million input tokens priced at just $0.15, while the output is priced at $0.60 per 1 million tokens. Also, with just 428 billion parameters, and 23 billion active ones, the model is currently leading globally on at least three widely used benchmarks - Automation Bench, Banker ToolBench, and LOCA-Bench-256K. This means that the model offers the absolute best performance-per-dollar for agent workflows. Even so, DeepSeek-V4.1-Flash-1017 still remains pareto on the DeepSWE v1.1 vs. cost per task curve, with MiniMax M3.1 Flash occupying the second place. The model offers five thinking levels - low, medium, high, xhigh and max (default) - and brings support for HTTP, Anthropic SDK, and OpenAI SDK. What's more, the MiniMax M3.1 Flash model brings an upgraded sparse attention framework as well as extensive architectural optimizations and a more refined hardware utilization setup. Per-token compute requirements reduced to 1/20th of the previous version An AI model basically consists of a series of transformer blocks that are made up of two key elements: - The Attention Layer looks at how words in a sentence connect with each other. For instance, if you give the model a prompt that says Paris is the capital of France, the attention layer will link the word 'capital' to 'France' and identify 'Paris' as the answer. Even so, this layer does not know what is meant by France or Paris, but would save these notes in the form of KV cache. As context increases, so does KV cache. - The Feed-Forward Network (FFN) holds the sum knowledge of a model in the form of weights, and it is this layer that uses deep mathematical formulas to unearth the meaning behind each word. This is typically known as activations or the hidden state, and constitute the internal mathematical representations that a model generates as tokens are processed by multiple transformer layers. For instance, it would look at France and activate its country encyclopedia to unearth the intended meaning of the word. MiniMax M3.1 Flash relies on its proprietary MiniMax Sparse Attention (MSA) framework to compensate for the KV cache bottleneck problem that can slow down prefill (input processing) and decode (output generation). Under the MSA framework, rather than evaluating attention on a token-by-token basis across the full 1-million-token context window, the KV cache is compressed into blocks. The system then calculates a low-cost "block score" to determine relevance. Only the most critical blocks (roughly a 6 percent to 7 percent slice) are actually pulled into full compute. This reduces per-token compute demands to a tiny 1/20th of the standard attention cost The Model speeds up prefill by 9x and decode by 15x MiniMax wrote custom GPU kernels that optimized how Query (Q) vectors fetch distributed KV cache data blocks described above. By restructuring the data access layout, the model achieves a 9x speedup during the initial prompt reading (prefill phase) and a 15x speedup during text generation (decode phase) at long contexts. Refined hardware utilization and memory efficiency Serving a Mixture of Experts (MoE) model with 428 billion total parameters across standard cloud clusters can lead to massive memory and networking bottlenecks. The MiniMax M3.1 Flash update focused on solving these engineering friction points: - Speculative Decoding Alignment: During the decode phase, the model feeds intermediate sequences through a faster, streamlined draft pathway to predict subsequent tokens, relying on the full 23-billion active parameter group to verify and correct them in parallel. Basically, a tiny draft model predicts upcoming tokens, while separate pathway verifies the prediction and corrects any prediction errors. - KV-Block-Major and Paged Attention: The architecture maps memory into dynamic page blocks using strided memory views, where the context is read using a strided pattern (for instance, attending to every N-th token or block), and then creates a virtual layer (or an illusion) of a condensed sequence without actually duplicating or allocating massive new physical memory pools on the GPU. This eliminates memory fragmentation entirely, keeping GPU utilization near maximum capacity without requiring massive, expensive memory overhead. - AB-Swapped HMMA Layout: The decoding mechanics feature customized matrix multiplication (HMMA) layouts specifically designed to track and score index branches rapidly without stalling GPU tensor cores. Native multimodality and system orchestration optimizations - Interleaved Pipeline Parallelism: Unlike models that tack a visual adapter onto a text-based model, the MiniMax M3.1 Flash uses a native visual encoder, and was trained on deeply interleaved text, image, and video data. The model routes visual and text-based tokens/signals through highly synchronized parallel pipelines, reducing data synchronization delays between different GPU nodes. - Rust-Based Serving Gateways: The visual and multimodal processing layer utilizes high-performance Rust-based Serving Model Gateways (SMG). This offloads heavy image/video preprocessing entirely to CPU clusters, ensuring that the GPU cluster is never waiting idle for data to be formatted. Because of all of these upgrades, MiniMax claims that the model offers: - Nearly 3× faster online inference - 70 percent lower cost for agentic tasks - 200+ TPS (tokens per second) output speed in everyday use (P50 170+) - Stronger coding, computer use, agentic and cybersecurity capabilities Both DeepSeek-V4.1-Flash-1017 and MiniMax M3.1 Flash illustrate China's growing compute constraints, even as the distillation debate remains inconclusive As is evident from the above X post, MiniMax currently has a compute footprint of just 70,000 H100-equivalent GPUs, while DeepSeek's is even lower, at just around 47,000 H100-equivalent GPUs. In comparison, OpenAI has a compute footprint of 3.7 million H100-equivalent GPUs! It is hardly a surprise, therefore, that the current models from both DeepSeek and MiniMax have fewer active and total parameters than their models that launched in 2024/2025, with China's AI labs increasingly forced to divert a greater share of their compute resources towards reinforcement learning (RL), where a model learns new tricks via a system of rewards and punishments. Given this paradigm, it is only natural for China's AI labs to resort to distilling the reasoning traces from the models of their Western counterparts, a task that has become much more difficult after the recent crackdown from Anthropic and OpenAI, with Anthropic going so far as to report a quarterly revenue loss of between 5 to 6 percent after it thwarted distillation pathways. And yet, the DeepSeek-V4.1-Flash-1017 model hallucinates a lot less frequently. If this model were distilled via Supervised Fine-Tuning (SFT), where a given LLM is trained on a curated dataset of high-quality, labeled input-output pairs (such as prompt/response or question/answer from a larger model), it would have hallucinated a lot more frequently, especially as this technique teaches the student model to answer with the confidence of the teacher model on facts it never stored. This is the enduring enigma: China's compute constraints necessitate distillation, and yet the effects of that distillation are not necessarily apparent in the output of some of the most successful models from China like the DeepSeek-V4.1-Flash-1017. Follow Wccftech on Google to get more of our news coverage in your feeds.