New AI models this week: Kimi K3.1 Pro and DeepSeek V4.1 Flash lead

Moonshot AI’s Kimi K3.1 Pro, listed on OpenRouter on October 23, is the most important release of the week from October 17 to October 24, 2026. DeepSeek and MiniMax added cheaper Flash models, and the open-weight side brought a time series model from Mistral, a memory model from Tencent and an image model from OpenMOSS.

Kimi K3.1 Pro: Moonshot’s long-context model with open weights promised

Moonshot AI says Kimi K3.1 Pro is the first model in a new K3.1 series and a post-trained version of Kimi K3. It is listed on OpenRouter with a context window of 1,048,576 tokens, at $2.00 per 1M input tokens and $10.00 per 1M output tokens. According to Moonshot, full open weights are due by November 4, and its own benchmarks put the model close to Claude Opus 5.5 and ahead of GPT-6 Astra on Terminal-Bench 4.0 (65.7%, against 66.4% and 58.2%). These are the maker’s figures and have not been independently checked. Details are in our earlier report on Kimi K3.1 Pro, and the OpenRouter listing has the API details, including tool use and reasoning controls.

DeepSeek V4.1 Flash 1017: A low-cost checkpoint with a new inference stack

DeepSeek’s V4.1 Flash 1017 is a Flash-tier checkpoint priced at $0.15 per 1M input tokens and $0.60 per 1M output tokens, with a 1,048,576-token context window. Its OpenRouter page lists reasoning controls, tool calling and log probabilities. DeepSeek also published DeepInfer, an Apache-2.0 inference framework that uses this checkpoint as its default. The README describes it as the framework behind DeepSeek’s large-scale production services. It supports expert parallelism, structured outputs, deterministic sampling and batch-invariant inference (see why batch invariance matters). It runs on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. The release does not include a KV cache system or prefill/decode disaggregation.

MiniMax M3.1 Flash: A cheap long-context option

MiniMax M3.1 Flash appeared on OpenRouter on October 19 with the same price and context window as DeepSeek’s Flash model: $0.15 per 1M input tokens, $0.60 per 1M output tokens and 1,048,576 tokens of context. The listing shows streaming, tool use, structured output and reasoning settings. It gives no architecture, license or benchmark details, so how it compares with DeepSeek’s model is unclear for now.

Timestral-209M: Mistral’s compact forecasting model

Mistral AI released Timestral, a family of time series foundation models for zero-shot forecasting in four sizes (5M, 28M, 123M and 209M parameters). The largest, Timestral-209M, is a patch-based transformer that handles up to 32,768 historical steps and predicts 21 quantiles. Its card claims state-of-the-art zero-shot accuracy on GIFT-Eval, TIME and fev-bench, including a MASE of 0.658 on GIFT-Eval. The license is a modified MIT.

Hy-Memory-30B-A3B: Tencent’s model for deciding what an assistant remembers

Tencent Hunyuan built Hy-Memory-30B-A3B as a memory “writer”: it decides what to keep from conversations and how new information updates a stored memory. It has 30B parameters with about 3B active per token, supports English and Chinese, and is released under Apache-2.0. It is meant to replace the general LLM inside Mem0, MemOS and Hy-Memory. On its card, it averages 55.49 across four harnesses, ahead of the listed Claude Opus 5 (54.62) and DeepSeek V4.1 Flash (53.64). The authors used fixed 20% subsamples of the benchmarks.

MOSS-Image: An open text-to-image model with layout control

OpenMOSS released MOSS-Image-Base and an SFT variant, with a ComfyUI version also on the Hub. The base model has about 5B parameters, uses an Apache-2.0 license and works in English and Chinese. Besides plain prompts, it accepts JSON layouts with bounding boxes and captions to control where text and objects appear. It needs an NVIDIA GPU, Python 3.10+ and PyTorch 2.9+.

Also seen this week

The community also picked up RecursiveSI/flint, XHToken/Spark-X2.6-4B and an uncensored FP8 build of DeepSeek V4.1 Flash 1017 from a third party. These are community uploads or lack an official page to check, so they are not covered here. A separate repository, openai/php2shell-checker, passed 1,001 GitHub stars in two days, but it is not a model.

The week in numbers

• Models with API prices: 3 (DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, Kimi K3.1 Pro), plus three open-weight releases on Hugging Face.
• Cheapest per 1M input tokens: $0.15, shared by DeepSeek V4.1 Flash 1017 and MiniMax M3.1 Flash.
• Most expensive per 1M input tokens: Kimi K3.1 Pro at $2.00 ($10.00 per 1M output tokens).
• Context window: 1,048,576 tokens for all three.

Prices for these and other models are tracked on our AI API pricing page.