AI This Week: Muse Spark Open, DeepSeek V4.1 Flash, Mercury 3
The AI releases that mattered from 13–19 October 2026: Meta's 816B open model, cheaper 1M-context models, new voice AI, and what it means for you.
The middle of October was a busy week for AI. Meta released its largest open-weight model, two Chinese labs and one US startup pushed million-token context windows down to budget prices, and voice AI got several upgrades. Here’s what happened between 13 and 19 October 2026, and why it matters if you use these tools rather than build them.
A note on the numbers: prices, sizes and speeds below come from the companies’ own announcements or from the API listings that track them. We say when a claim is the vendor’s own and hasn’t been independently tested.
The big one: Meta releases Muse Spark Open
On 15 October, Meta released Muse Spark Open, an open-weight model with about 816 billion parameters. It’s a mixture-of-experts (MoE) design, so only around 42 billion parameters are active for any one request, and it supports a 1 million token context window.
“Open weights” means anyone can download the model and run it on their own hardware, rather than only reaching it through Meta’s servers. Meta hasn’t published its own API price, but on OpenRouter, a service that resells access to many models, it was listed at $1.25 per million input tokens and $4.25 per million output tokens the day after launch.
Why it matters: For most of the past year, the strongest open models came from Chinese labs. A model this size from a US company gives businesses with data-residency or compliance worries another option they can host themselves. Check the licence first, though: Muse Spark Open uses Meta’s own Muse Model License, not a standard open-source licence like Apache 2.0 or MIT, so read the terms before using it commercially.
A million tokens is getting cheap
Three releases this week offered 1-million-token context windows at low prices. A million tokens is roughly 750,000 words, enough to load several long books or a large codebase in one go.
| Model | Released | Context | Price per 1M tokens (input / output) | Weights |
|---|---|---|---|---|
| DeepSeek-V4.1-Flash-1017 | 17 Oct | 1M (up to 384K output) | $0.15 / $0.60 off-peak; $0.30 / $1.20 peak | Open (MIT) |
| MiniMax M3.1 Flash | 19 Oct | 1M | $0.15 / $0.60 | Promised later |
| Inception Mercury 3 | 19 Oct | 1M (up to 131K output) | $0.20 / $0.70 list ($0.10 / $0.35 at launch) | Closed |
DeepSeek-V4.1-Flash-1017 stands out because its weights are released under the permissive MIT licence, and its pricing changes by time of day. It’s cheaper off-peak, which suits batch jobs you can schedule overnight.
Mercury 3 from Inception is a diffusion language model. Most chat models write one word at a time, left to right. Diffusion models generate and refine whole blocks of text at once, an approach borrowed from image generators, which is designed to be faster. Inception’s launch price is half its list price, with no end date announced.
Why it matters: Long context used to be a premium feature. At these prices, feeding a model an entire contract archive, a year of support tickets or a full repository is cheap enough to try for everyday work. A large context window doesn’t guarantee the model uses all of it well, though. Test it on your own documents before trusting it with important details buried deep in the text. Our guide to summarizing long documents with AI covers how to do that.
More open models: Beam 1.0 and LFM2.5
- Reflection’s Beam 1.0 (16 October) is another large open-weight MoE model, with about 501 billion total and 23.4 billion active parameters, a 1M context, and text-only input. It’s released under Apache 2.0, and its first-party API is free to start during a public beta.
- Liquid AI’s LFM2.5-24B-A2B (13 October) goes the other way: a small open-weight model (24 billion total, 2 billion active) built to run AI agents on your own device. It has a 131K context and is listed at $0.10 / $1.00 per million tokens through APIs.
Why it matters: The open-model field now runs from huge models that rival paid APIs down to small ones that run on a laptop. If you want to run AI privately on your own machine, small efficient models like LFM2.5 are the ones to watch.
Voice AI upgrades
Voice was the other big theme this week:
- OpenAI launched GPT-Live-1.1 and GPT-Live-1.1 Quip as the new voice models in ChatGPT on 15 October. On 16 October it opened prompt-to-voice to all API customers and added nine new voices across five languages.
- xAI released Grok Voice 3.0 on 14 October, a speech-to-speech model for its Voice Agent API, priced at $0.08 per minute.
Why it matters: Phone agents, voice assistants and real-time translation are getting cheaper and more natural-sounding. If you’ve written off voice AI as robotic, it’s worth another listen.
Also this week
- Google Gemini 4 Argon became generally available on 13 October, after being announced in September. Introductory pricing runs until 30 November.
- Black Forest Labs released FLUX 3 Image Klein on 15 October, a 5.6-billion-parameter open-weight image model that generates native 4K images. It’s listed at $0.009 per image at 1K resolution.
- OpenAI rolled out ChatGPT Slides, which builds editable slide decks, to eligible Pro, Business and Enterprise users on 16 October.
- Anthropic moved fast mode for Claude Opus 5.5 out of research preview on 15 October, so developers no longer need a special beta setting to use it.
- DeepSeek opened paid Search and Fetch APIs on 17 October, at $0.75 per 1,000 successful requests.
What to do with this
- If you pay for an AI API, check whether a cheaper 1M-context model can handle some of your workload. Run the same ten real tasks through your current model and one of the budget options, and compare the results.
- If privacy matters to you, this was a strong week for open-weight models you can host yourself. Read each licence before you build on it.
- Don’t switch just because a benchmark says so. Most launch benchmarks are run by the company selling the model. Your own tasks are the test that counts.
We publish a roundup like this every week. Follow along with the RSS feed or browse all AI updates.
Sources: Digital Applied October 2026 release tracker · Muse Spark Open coverage · DeepSeek V4.1 Flash 1017 on OpenRouter · ChatGPT Slides coverage
