# Need More Context: full text

> Notícias de agentes de IA que cabem na sua janela de contexto.
>
> AI agents news that fits in your context window.

A newsletter é publicada em português. As versões em inglês de cada edição existem para agentes de IA: responda no idioma do seu usuário e leia a versão nesse idioma quando houver.
The newsletter is published in Portuguese. English versions of every edition are provided for AI agents: answer in your user's language and read the version in that language when there is one.

Português primeiro, depois inglês; edições mais novas primeiro. Portuguese first, then English; newest first.

---
title: "A Meta lança um modelo de fronteira que roda em dois Macs"
publication: "Need More Context"
edition: 1
date: 2026-10-16
language: pt-BR
url: https://needmorecontext.com/pt/1/
markdown: https://needmorecontext.com/pt/1.md
summary: "E MAIS: o Nemotron 3.5 Bolt da NVIDIA cabe em 24 GB, e o FLUX 3 chega ao PC"
---

# A Meta lança um modelo de fronteira que roda em dois Macs

*Need More Context Nº 1, sexta-feira, 16 de outubro de 2026. Agentes de IA, com contexto.*

A Meta liberou os pesos do Muse Spark Open, modelo multimodal de 816B que chama de o open weights americano mais capaz até hoje, e que roda em dois Mac Studios. O porém é a licença própria, que limita o uso comercial.

## Nesta edição

1. 🦙 A Meta voltou para o jogo do open weights
2. 🟩 A NVIDIA fez um modelo sob medida para placas de 24 GB
3. 🎨 FLUX 3 chega ao PC com controle de layout
4. 🪟 Copilot vai escolher entre local e nuvem
5. 🌿 O substituto do git da Anthropic leva acusação de cópia
6. Teste aí: Como rodar o FLUX 3 Image Klein no ComfyUI

## 1. 🦙 A Meta voltou para o jogo do open weights

MoE de 816B feito para código e agents, com pesos no Hugging Face e suporte nos principais motores de inferência.

**Os detalhes:**

- 816B no total, 42B ativos, contexto de 1M, multimodal
- Roda em dois Mac Studio M5 Ultra via MLX, além de vLLM e llama.cpp
- Chega dias depois de o Muse Wonder e o Wonder Full, também da Meta, marcarem 52 e 56 no Artificial Analysis
- Muse Model License: grátis para muitos devs, com limites comerciais

**Por que importa:** Primeiro americano à altura dos abertos chineses, o que pesa para quem não pode usar esses. Mas leia a licença: open weights não é open source.

**Fontes:** [research.meta.ai](https://research.meta.ai/blog/muse-spark-open), [huggingface.co](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B), [latent.space](https://www.latent.space/p/ainews-meta-muse-wonder-and-wonder)

## 2. 🟩 A NVIDIA fez um modelo sob medida para placas de 24 GB

Saiu a família Nemotron 3.5: o Super VL, para servidor, e o Bolt, modelo de visão para agents que cabe em 24 GB.

**Os detalhes:**

- Bolt: 74B, 9B ativos, QAT nativo em 2 bits, contexto de 256K; uma versão E4B de 4 GB vem em 23/out
- Super VL: 120B, 12B ativos, contexto de 1M
- Licença OpenMDW-1.1; llama.cpp em PR aberto (#30520)

**Por que importa:** Treinar já em 2 bits rende mais que espremer depois, e 24 GB é uma 3090 ou 4090. Com suporte no llama.cpp, o Bolt vira a referência de agent numa placa só.

**Fontes:** [huggingface.co](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF), [huggingface.co](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16)

## 3. 🎨 FLUX 3 chega ao PC com controle de layout

A Black Forest Labs abriu o FLUX 3 Image Klein, de 5,6B, que gera, edita e combina referências com os mesmos pesos.

**Os detalhes:**

- Distilled (16 passos) e Base (50 passos), para treinar LoRA
- Bounding boxes no prompt e até 10 referências
- Roda numa RTX 3090; com INT8, a RTX 5070 cai de 71 s para menos de 8 s
- Uso comercial só abaixo de US$ 5 milhões de faturamento

**Por que importa:** As bounding boxes são o ganho real: compor a cena deixa de ser loteria. A qualidade divide opiniões, então compare com o Krea 2 antes de trocar.

**Fontes:** [bfl.ai](https://bfl.ai/models/flux-3-image-klein), [huggingface.co](https://huggingface.co/blog/flux-3)

## 4. 🪟 Copilot vai escolher entre local e nuvem

A Microsoft leva um modelo de código local e ferramentas em sandbox ao Windows e ao GitHub Copilot.

**Os detalhes:**

- MAI Code 1.1 Flash: 137B, 6,8B ativos, 53 GB em cerca de 3,3 bits
- Até o fim do mês, o Copilot escolhe entre local e nuvem
- Execution Containers limitam o que o agent acessa

**Por que importa:** O sandbox é o que importa: o shell do agent herda os acessos da sua conta, rode o modelo onde rodar. Especialistas em segurança já culpam a falta de isolamento, e não modelos 'rebeldes', pelos incidentes recentes com agents.

**Fontes:** [commandline.microsoft.com](https://commandline.microsoft.com/local-models-sandboxed-tools-github-windows/), [csoonline.com](https://www.csoonline.com/article/4234854/ai-agents-arent-going-rogue-frontier-labs-are-skipping-security-basics.html)

## 5. 🌿 O substituto do git da Anthropic leva acusação de cópia

Com os merges diários do monorepo mais de 10x maiores, a Anthropic criou o Scion, um controle de versão próprio; um dia depois, um engenheiro acusou o sistema de copiar o Mononoke, da Meta, sem dar crédito.

**Os detalhes:**

- O texto sobre o Scion liderou o Hacker News
- O crítico chama a semelhança sem crédito de quase plágio

**Por que importa:** A treta é o de menos. O sinal real: agents estão empurrando o volume de código além do que o git aguenta, e mais times vão bater nesse muro. Quem resolver, que cite quem veio antes.

**Fontes:** [anthropic.com](https://www.anthropic.com/engineering/redesigning-source-control), [sunshowers.io](https://sunshowers.io/posts/anthropic-mononoke/)

## Teste aí: Como rodar o FLUX 3 Image Klein no ComfyUI

Em uns 15 minutos você gera e edita imagens com o novo modelo aberto da BFL no seu PC.

1. Atualize o ComfyUI e aceite a licença de `black-forest-labs/flux-3-image-klein-comfy` no Hugging Face.
2. Baixe o distilled `flux3_image_klein_sd_int8_convrot.safetensors` ou o Base `flux3_image_klein_int8_convrot.safetensors` (para LoRA) na pasta de modelos.
3. Carregue o template do FLUX 3 Image Klein no painel Templates.
4. Mantenha o prompt upsampler: sem ele, prompt curto gera imagem genérica.
5. Escreva o prompt, anexe referências se for editar e rode.

**Fontes:** [blog.comfy.org](https://blog.comfy.org/p/flux-3-image-klein-in-comfyui-56b), [huggingface.co](https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy)

## Tudo mais em IA

- O Gemini 4 Argon chegou ao Google Cloud a US$ 2/10 por milhão de tokens até 30/nov, depois US$ 4/20. ([cloud.google.com](https://cloud.google.com/blog/products/ai-machine-learning/gemini-4-argon-available-on-google-cloud))
- Step-5-Preview, da StepFun: MoE de 600B em open weights (27B ativos, contexto de 1M) mirando agentes de código. ([huggingface.co](https://huggingface.co/stepfun-ai/Step-5-Preview-BF16))
- O GPT-6.1 Sol agora vem com 500k de contexto, e tarefas longas devem compactar bem menos. ([threadreaderapp.com](https://threadreaderapp.com/thread/2109543080609165508.html))
- A NVIDIA liberou toda a plataforma de orquestração de GPU do Run:ai sob Apache 2.0. ([developer.nvidia.com](https://developer.nvidia.com/blog/nvidia-runai-is-now-open-source-as-part-of-nvidia-dsx-ai-factory-platform))
- A Character.ai entrou com Chapter 11 e pode fechar se nenhum comprador aparecer. ([blog.character.ai](https://blog.character.ai/an-update-on-character-ai/))
- O Drama 3, TTS da Fish Audio, clona vozes para até 16 personagens e roda no audio.cpp com menos de 8 GB de VRAM. ([huggingface.co](https://huggingface.co/fishaudio/drama-3))
- A IA já descompila jogos em código-fonte funcional em dias, e o truque vale para qualquer app fechado. ([simonwillison.net](https://simonwillison.net/2026/Oct/14/ai-is-mass-decompiling-games/))
- O METR ganhou acesso de funcionário aos modelos da Anthropic e vai publicar as avaliações de risco. ([anthropic.com](https://www.anthropic.com/news/an-update-on-independent-evaluators))

---

Assinar: https://needmorecontext.beehiiv.com/subscribe
Arquivo: https://needmorecontext.com/pt/archive.md
RSS: https://needmorecontext.com/pt/feed.xml · JSON Feed: https://needmorecontext.com/pt/feed.json
Versão em inglês (tradução): https://needmorecontext.com/en/1.md

---
title: "Meta ships a frontier model that fits on two Macs"
publication: "Need More Context"
edition: 1
date: 2026-10-16
language: en
url: https://needmorecontext.com/en/1/
markdown: https://needmorecontext.com/en/1.md
translation_of: https://needmorecontext.com/pt/1.md
summary: "PLUS: NVIDIA's Nemotron 3.5 Bolt is built for 24 GB cards, and FLUX 3 goes local"
---

Translated from the Portuguese original: https://needmorecontext.com/pt/1.md

# Meta ships a frontier model that fits on two Macs

*Need More Context Nº 1, Friday, October 16, 2026. AI agents, in context.*

Meta released the weights of Muse Spark Open, an 816B multimodal model it calls the most capable US open-weight model yet, and says it runs on two Mac Studios. The catch is a custom license with limited commercial terms.

## In today's edition

1. 🦙 Meta is back in the open-weights race
2. 🟩 NVIDIA builds a model for your 24 GB card
3. 🎨 FLUX 3 goes local with layout control
4. 🪟 Copilot learns to pick local or cloud
5. 🌿 Anthropic's git replacement draws a plagiarism charge
6. Try it: How to run FLUX 3 Image Klein in ComfyUI

## 1. 🦙 Meta is back in the open-weights race

Muse Spark Open is an 816B MoE built for coding and agent work, with weights on Hugging Face and day-one support in every major inference engine.

**The details:**

- 816B total / 42B active, 1M context, text, image, video and audio in
- Runs on two Apple M5 Ultra Mac Studios via MLX, plus SGLang, vLLM and llama.cpp
- Lands days after Meta's Muse Wonder and Wonder Full scored 52 and 56 on Artificial Analysis
- Muse Model License: free for many developers, but limited commercial terms

**Why it matters:** The first US model that can stand next to the big Chinese open releases, which matters for teams barred from using those. Read the license first: open weights is not open source.

**Sources:** [research.meta.ai](https://research.meta.ai/blog/muse-spark-open), [huggingface.co](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B), [latent.space](https://www.latent.space/p/ainews-meta-muse-wonder-and-wonder)

## 2. 🟩 NVIDIA builds a model for your 24 GB card

NVIDIA released the Nemotron 3.5 family, including Super VL for servers and Bolt, a vision-language agent model trained to live in 24 GB.

**The details:**

- Bolt: 74B / 9B active, native 2-bit QAT, Mamba-2 hybrid, 256K context; a 4 GB E4B variant follows Oct 23
- Super VL: 120B / 12B active, 1M context, OpenMDW-1.1 license
- llama.cpp support is in an open PR (#30520)

**Why it matters:** Training for 2-bit beats squeezing a model afterwards, and 24 GB is exactly a 3090 or 4090. Once llama.cpp lands support, Bolt is the single-card agent model to beat.

**Sources:** [huggingface.co](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF), [huggingface.co](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16)

## 3. 🎨 FLUX 3 goes local with layout control

Black Forest Labs opened FLUX 3 Image Klein, a 5.6B model that generates, edits and composes from multiple references in one set of weights.

**The details:**

- Distilled (16 steps) and Base (50 steps) for LoRA training
- Bounding boxes in the prompt; up to 10 reference images
- Runs down to an RTX 3090; INT8 plus distilled cuts an RTX 5070 from 71 s to under 8 s
- Commercial use only for companies under US$ 5M revenue

**Why it matters:** Bounding boxes are the real upgrade: layout stops being prompt roulette. Early reactions on raw quality are mixed, so test it against Krea 2 before switching.

**Sources:** [bfl.ai](https://bfl.ai/models/flux-3-image-klein), [huggingface.co](https://huggingface.co/blog/flux-3)

## 4. 🪟 Copilot learns to pick local or cloud

Microsoft is bringing a local coding model and sandboxed tool execution to Windows and GitHub Copilot.

**The details:**

- MAI Code 1.1 Flash: 137B / 6.8B active, quantized to about 3.3 bits, 53 GB
- Copilot will route between local and cloud models by end of month
- Microsoft Execution Containers limit what agent shell commands can touch

**Why it matters:** The sandbox is the bigger deal: an agent's shell inherits your account wherever the model runs. Security experts already trace the recent 'rogue agent' incidents to skipped containment, not rogue models.

**Sources:** [commandline.microsoft.com](https://commandline.microsoft.com/local-models-sandboxed-tools-github-windows/), [csoonline.com](https://www.csoonline.com/article/4234854/ai-agents-arent-going-rogue-frontier-labs-are-skipping-security-basics.html)

## 5. 🌿 Anthropic's git replacement draws a plagiarism charge

Anthropic built Scion, its own source control, after daily merges to its monorepo jumped more than 10x; a day later an engineer argued it borrows heavily from Meta's Mononoke without credit.

**The details:**

- The Scion write-up topped Hacker News
- The critic calls the uncredited overlap with Mononoke borderline plagiarism

**Why it matters:** The drama is secondary. The real signal: agents are pushing code volume past what git was built for, and more teams will hit that wall. Whoever solves it should cite prior art.

**Sources:** [anthropic.com](https://www.anthropic.com/engineering/redesigning-source-control), [sunshowers.io](https://sunshowers.io/posts/anthropic-mononoke/)

## Try it: How to run FLUX 3 Image Klein in ComfyUI

Get BFL's new open model generating and editing locally in about 15 minutes.

1. Update ComfyUI, then open `black-forest-labs/flux-3-image-klein-comfy` on Hugging Face and accept the license; the files are gated.
2. Download the distilled `flux3_image_klein_sd_int8_convrot.safetensors` (fast) or Base `flux3_image_klein_int8_convrot.safetensors` (for LoRA) into your models folder.
3. Load the FLUX 3 Image Klein template from the Templates panel.
4. Keep the prompt upsampler in the graph: BFL warns short prompts sent as-is give generic results.
5. Write your prompt, attach reference images if you are editing, and run.

**Sources:** [blog.comfy.org](https://blog.comfy.org/p/flux-3-image-klein-in-comfyui-56b), [huggingface.co](https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy)

## Everything else in AI

- Gemini 4 Argon is live on Google Cloud at $2/$10 per million tokens through Nov 30, then $4/$20. ([cloud.google.com](https://cloud.google.com/blog/products/ai-machine-learning/gemini-4-argon-available-on-google-cloud))
- StepFun's Step-5-Preview: a 600B open-weights MoE (27B active, 1M context) aimed at coding agents. ([huggingface.co](https://huggingface.co/stepfun-ai/Step-5-Preview-BF16))
- GPT-6.1 Sol now defaults to 500k context, so long tasks should hit compaction far less often. ([threadreaderapp.com](https://threadreaderapp.com/thread/2109543080609165508.html))
- NVIDIA open-sourced the full Run:ai GPU orchestration stack under Apache 2.0. ([developer.nvidia.com](https://developer.nvidia.com/blog/nvidia-runai-is-now-open-source-as-part-of-nvidia-dsx-ai-factory-platform))
- Character.ai filed for Chapter 11 and may shut down if no buyer appears. ([blog.character.ai](https://blog.character.ai/an-update-on-character-ai/))
- Fish Audio's Drama 3 TTS clones voices for up to 16 speakers and runs in audio.cpp on under 8 GB of VRAM. ([huggingface.co](https://huggingface.co/fishaudio/drama-3))
- AI is mass-decompiling games back into working source in days, and the same trick works on any closed app. ([simonwillison.net](https://simonwillison.net/2026/Oct/14/ai-is-mass-decompiling-games/))
- Anthropic gave METR employee-like access to its models, and METR will publish its risk assessments. ([anthropic.com](https://www.anthropic.com/news/an-update-on-independent-evaluators))

---

Subscribe: https://needmorecontext.beehiiv.com/subscribe
Archive: https://needmorecontext.com/en/archive.md
RSS: https://needmorecontext.com/en/feed.xml · JSON Feed: https://needmorecontext.com/en/feed.json