DeepSeek and MiniMax launch low-cost, high-performance AI models

DeepSeek MiniMax AI agents large language models DeepInfer

Chinese AI developers DeepSeek and MiniMax have released new high-speed, low-cost models targeting the AI agent and coding markets.

MiniMax launched M3.1 Flash, which was previously known under the anonymous pseudonym ‘Space Bunny’. The model utilizes a mixture-of-experts (MoE) architecture with 428 billion total parameters. MiniMax claims the model reduces AI agent task costs by 70% compared to its predecessor and offers inference speeds nearly three times faster, reaching over 200 tokens per second. It features a one-million-token context window and supports text, image, and video inputs.

DeepSeek released the V4.1-Flash-1017 model, focusing on logical reasoning, coding, and visual recognition. DeepSeek also open-sourced DeepInfer, an Apache 2.0 licensed inference framework designed to optimize the performance of its models on NVIDIA and Huawei hardware. DeepInfer includes features such as speculative decoding and distributed parallelism to increase throughput and reduce latency.

Both companies are employing aggressive pricing strategies to compete with Western providers. MiniMax M3.1 Flash is priced at $0.15 per million input tokens and $0.60 per million output tokens, while DeepSeek’s V4.1-Flash offers competitive rates aimed at maximizing cost-efficiency for developers.

Entities

Artificial Analysis · DeepInfer · DeepSeek · GLM · Google · Huawei · MiniMax · Nvidia

Claims

What the coverage asserts, and how many sources carry each claim.

[● 4 SOURCES] MiniMax M3.1 Flash is priced at $0.15 per million input tokens. coinacademy.fr · weel.co.jp · siliconangle.com · gigazine.net
[● 5 SOURCES] MiniMax M3.1 Flash has 428 billion total parameters. coinacademy.fr · weel.co.jp · siliconangle.com · gigazine.net
[● 3 SOURCES] DeepInfer is optimized for NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. cryptobriefing.com · www.infoq.cn · dev.to
[● 4 SOURCES] MiniMax M3.1 Flash reduces AI agent task costs by 70% compared to previous models. coinacademy.fr · weel.co.jp · siliconangle.com · gigazine.net
[● 2 SOURCES] DeepSeek V4.1-Flash-1017 resolved 77% of tasks on the DeepSWE v1.1 test. technews.tw · siliconangle.com
[● 4 SOURCES] MiniMax M3.1 Flash features a context window of one million tokens. coinacademy.fr · weel.co.jp · siliconangle.com · gigazine.net
[● 3 SOURCES] MiniMax M3.1 Flash achieved a score of 76.2 on the BankerToolBench evaluation. weel.co.jp · siliconangle.com · gigazine.net
[● 4 SOURCES] DeepSeek open-sourced the DeepInfer inference framework under the Apache 2.0 license. the-decoder.de · cryptobriefing.com · www.infoq.cn · dev.to