AI Model Launch Tracker

What we captured

MiniMax announced an updated M3.1 Flash model and claimed improvements in agentic tasks, programming, cybersecurity, inference speed, and usage costs.

• • MiniMax said it launched the updated model on October 19, 2026.
• • The company claimed nearly triple online inference speed and a 70% reduction in usage costs.

Source evidence

Confidence 99% / Completeness 79%

What we captured

DeepSeek released V4.1-Flash-1017, an upgraded version of V4.1-Flash, with claimed improvements in speed and several capabilities.

• • DeepSeek released V4.1-Flash-1017 as an upgrade to V4.1-Flash.
• • DeepSeek listed output-token API prices of $0.59 per million off-peak and $1.19 per million at peak hours.
• • DeepSeek's own DeepSWE v1.1 evaluation gave the model a score of 77.0.

Source evidence

Confidence 96% / Completeness 79%

What we captured

Together AI lists DeepSeek V4.1 Flash 1017 as an API-available update to DeepSeek V4.1 Flash, with a 1M-token context window and MIT license.

• • Together AI describes the model as an update to DeepSeek V4.1 Flash.
• • The model accepts text and image input and has a 1M-token context window.
• • The model card states that 8B parameters activate during prefill and 16B during decode, from a 552B backbone.
• • Together AI lists October 16, 2026 as the release date.

Source evidence

What we captured

Reflection AI released the weights of Beam, its first open model, under the Apache 2.0 license.

• • Beam is Reflection AI’s first open model.
• • The model has 501 billion total parameters and 23 billion active parameters per token.
• • Reflection says Beam was trained from scratch using four weeks of reinforcement learning.
• • Beam’s released weights use the Apache 2.0 license.

Source evidence

What we captured

Meta released Muse Spark as an open-weight model for coding and AI agents, with downloads and access through select API providers.

• • Meta announced the release on October 16, 2026.
• • Meta reported 816 billion total parameters and 42 billion active parameters.
• • The model supports a context window of up to 1 million tokens.
• • The article reports local operation on two Apple M5 Ultra Mac Studio computers.

Source evidence

What we captured

Meta released Muse Spark Open, an 816-billion-parameter open-weight model under its custom Muse Model License.

• • Meta announced the release on October 15, 2026.
• • The model has 816 billion parameters and activates 42 billion per token.
• • Together AI lists serverless rates of $1.25 per million input tokens and $4.25 per million output tokens; these are provider-specific rates.
• • The model is released under Meta’s custom Muse Model License.

Source evidence

What we captured

GPT-Live-1.1 is rolling out to all users of ChatGPT, Codex, and dots, with 13 new voices and reported improvements to multilingual performance, delegation, and steerability. API availability is planned for later.

• • The post is dated October 15, 2026, and says the rollout is happening that day.
• • The update adds 13 voices and reports improved performance in several non-English languages.
• • The API is not yet available according to the announcement.

Source evidence

What we captured

Liquid AI released d1-ultra, a multimodal decision model that returns probabilities for text and image inputs.

• • Released October 15, 2026.
• • Accepts text, images, or both and returns probabilities for defined decisions.
• • Does not generate or bill for output tokens.
• • Input price: $0.042 per million tokens.
• • Liquid AI says it leads the company's 112-test suite.

Source evidence

Confidence 99% / Completeness 92%

What we captured

Skywork AI released Matrix-Game 4.0, an open-source interactive game-generation system combining an executable engine with a causal video model.

• • Released October 15, 2026 and described as open source.
• • Pairs an executable game engine with a causal video model for interactive rendering.
• • Skywork claims 480p at 60 FPS on one NVIDIA RTX 5090 and 112 ms from key press to first responding frame.
• • The performance claims are not independently verified and are limited to scene-specific setups.

Source evidence

Confidence 95% / Completeness 87%

What we captured

Multiverse Computing released downloadable weights for Comet 60B, a compressed version of NVIDIA’s Nemotron 3.5 Super VL that retains text, image, and video input.

• • Released October 15, 2026.
• • Compressed version of NVIDIA’s Nemotron 3.5 Super VL, with roughly half its total parameter count.
• • Retains text, image, and video input; weights are downloadable in three formats.
• • Accuracy and throughput comparisons are company-reported.

Source evidence

Confidence 99% / Completeness 87%