New AI models this week: Kimi K3.1 Pro and MiniMax M3.1 Flash lead

Two Chinese labs shipped the week’s main models: Moonshot AI released Kimi K3.1 Pro on October 23, and MiniMax followed on October 19 with M3.1 Flash. Both offer a context window of about 1 million tokens, but their input prices differ by more than tenfold. On Hugging Face, a coding-agent model from Recursive, a small on-device model from SparkLLM, and specialized releases from Mistral, Tencent and the OpenMOSS team added to the list.

Prices below are in US dollars per 1M tokens and come from our price tracker, which reads OpenRouter’s catalogue. The full list is on our AI API pricing tracker.

Kimi K3.1 Pro: Moonshot’s long-task model with a trimmed cyber skill

Moonshot AI describes Kimi K3.1 Pro as a post-trained version of Kimi K3 and its most capable 3T-class open-weight model. It targets coding, long-horizon agent tasks, 3D modeling and CAD, and financial workflows. The context window is 1M tokens, and a Highspeed variant, which Moonshot says runs 2 to 3 times faster, is offered alongside it.

On OpenRouter the model costs $2.00 per 1M input tokens and $10.00 per 1M output tokens. Moonshot reports 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5 in the same table, and says its hallucination rate on a hard stress test fell from 23.4% to 8.3%. These are company-reported figures and have not been independently verified.

Moonshot says the full weights will be released by November 4, and that it removed cyber-related data from training for that open release. We covered the cyber trimming and the OpenRouter listing in separate articles.

MiniMax M3.1 Flash: A low-priced multimodal model for agents

MiniMax’s M3.1 Flash is a native multimodal mixture-of-experts model that takes text, image and video as input and returns text. Its model card lists about 428B total parameters with 23B active per token, and aims it at coding, agentic workflows, computer use and cybersecurity. Reasoning effort is adjustable from low to max.

The card reports 73.7% on DeepSWE v1.1 and 61.7 on OSWorld 2.0, and notes that DeepSeek V4.1 Flash-1017 scores higher on DeepSWE, at 77.0. The license is listed as “minimax-community.”

The price is the notable part: $0.15 per 1M input tokens and $0.60 per 1M output tokens, with a 1,048,576-token context window. That makes it the cheapest model on the tracker this week, at less than a tenth of Kimi K3.1 Pro’s input price.

Flint: Recursive’s open coding-agent model that runs on one GPU

Recursive Superintelligence published Flint, a mixture-of-experts language model with 39.8B total parameters and about 2.5B active per token. Its card says it is designed to investigate codebases, fix bugs, implement features and write tests, with reasoning and multi-turn tool use.

The company reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified from its release evaluations. These are self-reported and not independently checked. The weights use the Apache 2.0 license, while hosted API terms are separate. The BF16 checkpoint is about 80 GB and is meant to run on a single B300 or H200 GPU per replica with the included SGLang kit. The documented context is 262,144 tokens, and Transformers modeling code is not yet included.

Spark-X2.6-4B: A 4B on-device model with a 1M-token setting

SparkLLM’s Spark-X2.6-4B is a compact general-purpose model for on-device use, aimed at agentic tasks, coding, long-context work and reasoning in English and Chinese. It keeps the architecture of Spark-X2.5-4B, and the team attributes the gains to improved post-training. Thinking has three levels (none, low and high) and is on by default.

The card’s SGLang example sets a maximum context of 1,048,576 tokens, which it says needs enough device memory. The authors claim best-in-class results among models of similar size, but the text gives no numeric scores, so those claims cannot be checked from the page. The license is Apache 2.0.

Timestral-209M: Mistral’s small time series forecaster

Mistral AI’s Timestral-209M is a time series foundation model for zero-shot forecasting. It is the largest of four sizes (5M, 28M, 123M and 209M parameters). It handles univariate and multivariate series, covariates and probabilistic forecasts across 21 quantiles, taking up to 32,768 past time steps and forecasting up to 1,024 steps at once. The license is a modified MIT license, and the page says a tech report is coming soon.

Hy-Memory-30B-A3B: Tencent’s model for writing agent memory

Tencent Hunyuan’s Hy-Memory-30B-A3B has 30B parameters with about 3B active per token. It is not a general chatbot: it acts as the write-path model inside external memory frameworks such as Mem0, MemOS and Tencent’s own Hy-Memory, deciding what to keep and how new information updates the store. The weights use the Apache-2.0 license.

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

OpenMOSS’s MOSS-Image-Base is a text-to-image model of about 5B parameters under the Apache 2.0 license. Besides plain prompts it accepts tag sequences and JSON layouts that control where text and objects appear. A fine-tuned MOSS-Image-SFT variant is also published, and Comfy-Org lists a packaged version for ComfyUI.

DeepInfer: DeepSeek opens its inference framework

Not a model, but relevant to anyone running one: DeepSeek published DeepInfer, which its page describes as the framework behind DeepSeek’s large-scale production services. The open release covers offline inference for DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. It lists expert parallelism, speculative decoding, asynchronous scheduling and structured outputs, and is licensed under Apache 2.0. A KV cache system and prefill/decode disaggregation are not included in this release. The repository passed 2,066 stars on GitHub within four days.

The week in numbers

• Models on our price tracker: 2 (MiniMax M3.1 Flash and Kimi K3.1 Pro), plus four notable Hugging Face releases covered above.
• Cheapest per 1M input tokens: MiniMax M3.1 Flash at $0.15.
• Most expensive per 1M input tokens: Kimi K3.1 Pro at $2.00.
• Context window: 1,048,576 tokens for both priced models.