This Week's Releases: Meta Opens Muse Spark, Reflection Ships Beam, GPT-6 Pro Gets Parallel Agents and Google Builds Gemini for Government
Title: This Week's Releases: Meta Opens Muse Spark, Reflection Ships Beam, GPT-6 Pro Gets Parallel Agents and Google Builds Gemini for Government
Date: 21 October 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1564 words
Reading time: 6 min
Published: 21-10-2026
Meta released the weights of Muse Spark Open on 15 October. Meta says the 816-billion-parameter model can run on two Apple Mac Studios. Reflection AI published the weights of its 501-billion-parameter Beam model on 16 October, under the Apache 2.0 licence. OpenAI gave GPT-6 Pro in ChatGPT the ability to split a task across several agents working at once, also on 16 October. DeepSeek updated its V4.1 Flash model on 17 October and opened a paid search API. Microsoft announced new controls for AI agents inside companies on 19 October. Google Public Sector put Gemini for government services into preview on 20 October.
Meta's Muse Spark Open
Meta Superintelligence Labs released Muse Spark Open on Thursday 15 October. It is a mixture-of-experts model with 816 billion parameters, of which 42 billion work on each token. The context window is one million tokens. Meta calls it the most powerful American open-weight model to date. Muse Spark, released in April, was Meta's first model without public weights.
The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. They come under Meta's own Muse Model License, which Meta says has "limited commercial terms, which allow many developers to use the model for free". The licence sets two conditions, according to reports. Products built on it must display "Built with Muse Spark Open". And a company that sells the model's outputs as a service needs a separate written agreement with Meta once its gross revenue passes $20m in any twelve months. Coverage describes it as open-weight rather than open source.
Meta and Apple say the model runs on two M5 Ultra Mac Studios linked together, using Apple's MLX software. Reports disagree on how much memory the 4-bit version needs, with figures from about 493GB to about 700GB. CNBC put the list price of one 256GB M5 Ultra Mac Studio at $10,799 or more. Meta has not published hosted token prices.
Most benchmark figures so far are Meta's own, and a detailed evaluation report is still to come. Meta's comparisons put it ahead of Kimi K3, GLM 5.3 and Qwen3.8 on agent and tool-use tests. It trails on coding: on Terminal-Bench 2.1 it scores 76.4, against 85.0 for Kimi K3 and 83.9 for GLM 5.3. Published Artificial Analysis data gives it 39 on the Intelligence Index, one point above Mistral Large 4.
Reflection's Beam
Reflection AI, a US company backed by Nvidia, released the weights of Beam on Hugging Face on Friday 16 October. Reflection first showed Beam on 5 October with a waiting list. Beam is text only. It has 501 billion parameters, about 23 billion of them active per token, and a context window of one million tokens. It is aimed at coding, reasoning and agent work.
Beam uses the Apache 2.0 licence, which allows commercial use without a revenue threshold. Reflection published the weights in three formats: BF16, FP8 and the smaller NVFP4. It also released a separate drafter model that speeds up generation. The NVFP4 version is 273 GiB, and Reflection's serving guides assume four Nvidia B300 graphics processors. Reflection published a model card and a safety technical report. It says more than 40 partners are involved, without a full list.
Reflection's own API is in public beta and free for now. Hosted prices elsewhere vary widely. Nebius, through OpenRouter, lists $0.25 per million input tokens and $1.00 per million output tokens. Together AI lists $1.40 and $4.40 for its reasoning route.
Reflection's model card reports 80.9 per cent on SWE-bench Verified and 90.5 per cent on GPQA Diamond. One outlet gives 81.6 per cent for the BF16 version. Artificial Analysis independently scored Beam 30 on its Intelligence Index, second among US open models.
Parallel agents in GPT-6 Pro
OpenAI updated GPT-6 Pro in ChatGPT on Friday 16 October. OpenAI says the model now answers faster and makes better use of plugins and tools. It can also split one request into several parts and hand each to a separate agent working at the same time. Users can follow each agent's progress in the conversation and open a finished task to see what that agent did.
The feature is rolling out gradually to eligible Pro users in the Chat tab. In managed Business and Enterprise workspaces, the workspace settings decide access. GPT-6 Pro is not part of the Plus plan. OpenAI has not said whether parallel agents will reach the API, and its API changelog for 16 October lists voice updates only.
OpenAI has not published speed or quality figures, a limit on the number of agents, or how multi-agent work counts against a Pro user's allowance. A small test by Kingy.ai found the parallel runs took about 9 per cent less time on average than sequential ones, with no clear gain in accuracy. Kingy.ai says four runs are not a benchmark.
On Saturday 17 October OpenAI also changed how Codex uses its allowance. Subagents can now reuse context, cached context stays available for longer, and requests are routed to raise the share that hits the cache. OpenAI says the aim is to give users more effective usage from the same plan.
DeepSeek's V4.1 Flash update and search API
DeepSeek released V4.1-Flash-1017 on Saturday 17 October. It is a dated update to V4.1 Flash, not a new model family. The weights are on Hugging Face and ModelScope under the MIT licence. Developers calling the model by its existing API name get the update without any change to price.
DeepSeek's own evaluation puts it at 77.0 on DeepSWE v1.1, a coding test, up from 74.2 for the original V4.1 Flash. The model accepts text and images and has a one-million-token context window. Reports disagree on the size of the model behind it, with figures of 552 billion and 763 billion parameters.
DeepSeek also opened two paid tools, a Search API for web searches and a Fetch API that retrieves the content of a page. Each costs $0.75 per 1,000 calls. Server-side searches made through DeepSeek Harness, Claude Code or Codex are free for API users until 31 October at 16:00 UTC, when charging begins.
Microsoft's controls for agents at work
Microsoft extended Project Perception, its security programme for AI agents, on Monday 19 October. Hayete Gallot, a Microsoft security executive, said frontier models "can't be trusted", from whichever company. She said the reason is not malice: models trained to get results can find shortcuts, and they are not tuned to an organisation's own rules.
The main addition is Guardian Agents, which watch the actions of other agents and carry out playbooks that security staff set up. Microsoft Defender can now detect 60 types of agent running on company machines, up from 25, including Claude Code, Codex and GitHub Copilot. Each agent it finds is given an identity in Microsoft Entra Agent ID, separate from the person who runs it.
Microsoft also described three proposed controls: Dynamic Permissions, Dynamic Agent Risk and Dynamic Conditional Access. They would limit each agent to the access its current task needs and change that access as the risk changes. SiliconANGLE reports that Microsoft gave no availability dates or prices. The controls sit within Project Perception and Microsoft 365 E7.
Gemini for government services
Google Public Sector launched Gemini for government services in preview on Tuesday 20 October, for US federal, state and local agencies. It is a version of the Gemini agent Google Cloud announced on 8 October, with parts built for public-sector work. The US Department of Health and Human Services and the State of Missouri are early users.
It has three parts. Skills are instructions written by subject experts that apply the statutes, regulations and agency guidance behind a task. Freedom of Information Act compliance and redaction is one of the built-in skills. Connectors link the agent to case management, finance, HR and collaboration systems, and Google says the agent queries each system where it sits rather than copying the data. Partner agents handle jobs such as finding records for an information request and suggesting redactions in batches.
Google says a person approves each consequential step, and answers can cite the source documents. There is no public price or date for general availability.
Before switching
Most of the benchmark figures this week come from the companies that built the models. Artificial Analysis's scores for Muse Spark Open and Beam are the main independent numbers.
The two big open-weight releases come with different terms. Beam's Apache 2.0 licence has no revenue threshold. Muse Spark Open needs a separate agreement above $20m of revenue from selling its outputs, and products built on it must carry Meta's attribution line. Legal teams will want to read the Muse Model License before a pilot becomes a product.
Running either model in house needs hundreds of gigabytes of memory, so a hosted route is the realistic first test for most organisations. Hosted prices for Beam already differ several times over between providers.
In SAASiQ's view, Microsoft's and Google's announcements point the same way: agents get their own identities and a person signs off the steps that matter. That gives identity and audit teams something they can review like any other account.
DeepSeek starts charging for server-side searches through Claude Code and Codex on 31 October. Microsoft has not given dates for its new agent controls, and Google has not given a general availability date for Gemini for government services.
SAASiQ - Intelligent Solutions for SaaS ©
