The brief
Five things that moved. What they mean. What to ignore.
01 · Open / local · 10/15/26 · Meta
Meta opens 816B weights — and says two Mac Studios can run them.
Thursday, Meta Superintelligence Labs released Muse Spark Open: an 816B-parameter mixture-of-experts with 42B active, natively multimodal (text, image, video, and audio in; text out), and a 1M-token context. Meta pitches it for coding and long-horizon agents, and says it is the first in its class to run on a desktop via two Apple M5 Ultra Mac Studios with MLX. Day-one stacks also include SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp, and Unsloth, with hosting on Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together. Weights sit on Hugging Face as meta-models/Muse-Spark-Open-816B-A42B. Meta’s own table puts it at AA Intelligence Index 39, JobBench 61.8, and MCP-Atlas 89.2. The license is the Muse Model License — free for many uses, with “Built with Muse Spark Open” attribution — and if you offer Model-as-a-Service and affiliate revenue exceeds $20M in any 12 months, you must notify Meta and get a separate written deal before offering it as Maas. Catch: it is not Apache or MIT; the desktop path is two M5 Ultra Mac Studios, not a laptop; the GPU path is huge (≥700GB NVFP4, ≥2TB BF16 on the HF card); and the benches are Meta’s.
Meta, 10/15/26. Meta developer blog. Muse-Spark-Open-816B-A42B.
02 · Open / local · 10/15/26 · Nvidia
Nvidia’s answer runs from the rack to your desk.
The same day, Nvidia released Nemotron 3.5 Super VL — a 120B-total / 12B-active multimodal MoE with up to 1M context for text, image, document, and video — plus Nemotron Bolt 74B-A9B quantized natively to 2-bit QAD so it targets 24GB. A smaller Bolt E4B (4GB) is due 10/23/26. Nvidia claims up to 5x higher throughput versus other open models in class on a single GPU, and Artificial Analysis’s Intelligence Index sits around 26 (the U.S. open lead in the AA write-up Nvidia points to). Open weights, datasets, and NeMo recipes ship with day-one paths through Hugging Face, OpenRouter, ModelScope, and NIM; Bolt is aimed at llama.cpp, Ollama, LM Studio, Unsloth, and vLLM on RTX and DGX Spark gear, with RTX Spark laptops arriving 10/16/26. Catch: the speed and accuracy claims are Nvidia’s and partner benches; Bolt E4B is not out yet; and Super VL is still a big model even at 12B active.
03 · Product · 10/14/26 · xAI / SpaceXSI
Grok Voice 3.0 tops the speech-to-speech board at $0.08 a minute.
Wednesday, xAI shipped Grok Voice 3.0 on the Voice Agent API. The grok-voice-latest pointer already lands on 3.0, and Voice 2.0 is deprecated. Price is unchanged: $0.08 per minute ($4.80 per hour). Artificial Analysis’s Speech-to-Speech Index, cited in the post from a 10/13/26 source, puts Voice 3.0 at 86.9 — #1 ahead of GPT-Live-1 Astra at 83.2 and Gemini 3.8 Live at 82.6 — with a Full Duplex Bench of 97% and τ-Voice at 71.6% at about 40% lower latency than GPT-Live-1 on that test (company framing). xAI says the stack already runs in Tesla vehicles and Starlink call centers. Catch: the AA numbers are independent scores cited by xAI; the Starlink and Tesla operational figures are company-reported.
04 · Money · 10/14/26 · Anthropic
Anthropic tells investors Q3 revenue hit $16.8B.
Anthropic told prospective investors that third-quarter revenue hit $16.8B — up about 45% from Q2 — and said it was profitable on an adjusted basis and cash-flow positive. IPO talk is still November-ish. Tuesday’s letter flagged the Wednesday investor meeting; this is the figure that came out of it. Catch: it is sourced reporting, not a public filing, and how partner revenue is counted remains murkier than a clean apples-to-apples against OpenAI’s earlier investor number.
05 · Open · 10/15/26 · StepFun
StepFun drops a 600B preview with 27B active.
Thursday, StepFun put Step 5 Preview on Hugging Face: a 600B-total / 27B-active sparse MoE with up to 1M context, text and image in / text out, aimed at agentic coding and knowledge work with a finance tilt. BF16, FP8, and NVFP4 builds ship with vLLM and SGLang day one at stepfun-ai/Step-5-Preview. The license is StepFun Model License 1.0 — Apache-2.0-style with extra attribution and commercial conditions, not plain Apache. Catch: it is not Apache; you need serious multi-GPU; and the name still says Preview.
The tape
Same-window notes that did not need their own brief slot. Named. Dated. Written so you can follow them without leaving the page.
Model · 10/14/26 · Paper Instruments
Ultramarine 1 Flash is API-first — weights in about two weeks.
A five-person lab post-trained GLM-5.3 Flash with non-verifiable RL for finance, law, consulting, and ops. The API is live as ultramarine-1-flash; weights are promised in about two weeks. Paper Instruments claims +4.1 points average versus the base on six knowledge-work benches and $0.13 per task on its internal suite versus much costlier frontier APIs on the same suite. Catch: most comparisons are Paper Instruments’ own suite and harness, and the weights are not out yet.
Open · 10/15/26 · IFM / Inception42
Jais 3 opens a six-model Arabic fleet.
IFM and Inception42 released Jais 3: six open Arabic models from 0.9B to 375B-A23B, plus Jais Speech, built on K2 Horizon with 512K context, on Hugging Face with vLLM and SGLang. Catch: it is Arabic-focused, and the Arabic eval jumps are company-reported.
Hardware · arriving 10/16/26 · Nvidia
RTX Spark laptops land Friday for local Bolt agents.
Nvidia’s Nemotron Bolt post flags 10/16/26 as the arrival date for RTX Spark laptops meant to run local Bolt agents on the 24GB path above. Catch: arrival date is Nvidia’s; availability and configs will vary by vendor.
The stack
What you can actually open. One desktop-class open multimodal, one local agent aimed at 24GB (32GB in practice), one multi-GPU preview from this window.
Model · 10/15/26 · Meta · Muse Model License
Muse Spark Open — two Mac Studios, or a serious GPU rack.
Grab meta-models/Muse-Spark-Open-816B-A42B on Hugging Face. Meta’s desktop story is two M5 Ultra Mac Studios (512GB each) via MLX; the GPU card lists ≥700GB NVFP4 and ≥2TB BF16. Hosted endpoints exist if you do not want the iron. Catch: Muse Model License, not Apache — watch the $20M Maas trigger and the “Built with Muse Spark Open” line.
Model · 10/15/26 · Nvidia · open weights
Nemotron Bolt 74B aims at 24GB — llama.cpp, Ollama, LM Studio.
Bolt 74B-A9B is the openable piece of the Nemotron 3.5 drop if you want a local agent on a single 24GB card. Day-one tooling includes llama.cpp, Ollama, LM Studio, Unsloth, and vLLM. Heads up: it uses a brand-new hybrid architecture, so it needs llama.cpp v0.7.0 or newer (out 10/15/26); older builds won't load it. Despite the name, NVIDIA's own setups ask for a 32GB card or bigger; on 24GB expect text-only with a short context. Catch: Super VL is the bigger multimodal sibling; Bolt E4B (4GB) is still due 10/23/26.
Model · 10/15/26 · StepFun · StepFun Model License 1.0
Step-5-Preview — multi-GPU FP8/BF16, not a laptop toy.
Open stepfun-ai/Step-5-Preview for the 600B / 27B-active sparse MoE. vLLM and SGLang work day one; plan for multi-GPU FP8 or BF16. Catch: StepFun Model License 1.0 adds attribution and commercial conditions beyond plain Apache.
Tool · 10/15/26 · llama.cpp v0.7.0 · MIT
llama.cpp v0.7.0 turns on MTP for GLM-5.3-Flash — plus LiliCorr drafts.
GLM-5.3-Flash (Z.ai’s 320B-total / 18B-active GLM-5 model) already loaded in llama.cpp as of 9/30; v0.7.0 adds its multi-token prediction head, so the model drafts its own next tokens. Turn it on with --spec-type draft-mtp (testers used --spec-draft-n-max 2), using a GGUF with the MTP layer built in or the separate mtp- file from ggml-org/GLM-5.3-Flash-GGUF. Testers in the PR report about 2x faster decode at short context and +52% at 32K, and 11.7→18.4 tok/s on 5x MI50. The same release adds LiliCorr, a new speculative-decoding method that picks draft candidates with a small draft transformer on top of DFlash. Catch: this is still a 320B model — even a 1-bit quant is a tight fit on a 96GB RTX PRO 6000, and MTP’s draft context costs VRAM (one tester dropped from 128K to 48K context). Single-card setups need big system RAM with experts offloaded. Older Unsloth GGUFs use the old glm5next name and need their Shard_Rewrite files.
llama.cpp v0.7.0 release. PR #29928 (GLM5-Next MTP). PR #30506 (LiliCorr). zai-org/GLM-5.3-Flash.
AI Gigs
AI companies are hiring, and not just coders. These six are open now, and you can apply today. Where pay is listed, it's what the company posted, and every link goes straight to the job.
Full-time · US remote · 10/08/26 · Robots and Pencils
AI Engineer
Robots and Pencils · $105K–$147K a year. Early-career role building and shipping AI/ML systems for client products.
Full-time · US remote · 10/09/26 · fal
Senior Software Engineer, ML Infrastructure & Automation
fal · $170K–$230K a year + equity. Infra for the generative-media inference platform: ML systems, automation, scale.
Contract · North America remote · found 10/15/26 · Turing
Senior Software Engineer, AI Evaluation / Coding Agents
Turing · $100–$150 an hour. About 3 months, 20–40 hrs/week: review coding agents’ work in real repos and write the rubrics.
Contract · remote worldwide · found 10/16/26 · G2i
Python Senior Software Engineer, AI Training
G2i · $100–$200 an hour. About 10–20 hrs/week: evaluate coding-agent interactions (Codex / Claude Code) and write clear feedback. Not a production coding seat.
Contract · remote / Palo Alto · found 10/15/26 · Elorian
Model Safety Contractor
Elorian · $75–$125 an hour. About 40 hrs/week: run model safety evals and curate safety data for a multimodal AI lab.
Contract · US remote · found 10/16/26 · Hippocratic AI
RN Clinical Data Annotation Specialist
Hippocratic AI · Pay not listed. About 20–40 hrs/week: score their AI nurse's conversations. Needs an RN license and 5+ years of clinical experience.
Grok Bot corner
New here? Start with this.
Meet Grok Bot
A helper that works for you, even when you're not looking.
Grok Bot is an AI assistant with its own computer. You tell it what you need in plain words, and it does the job for you, like planning your day, looking something up, or writing a note. It remembers what you like, and it can do things on a schedule, so a reminder or a morning summary just shows up when you want it.
Lesson 1 · Talk, don't type
Take Grok Bot on your walk.
You don't have to type. Start a voice call and just talk, like a phone call with a friend who helps. When you hang up, you get a short summary of what you talked about.
Try it: start a voice call and say, "I'm out for a walk. Help me plan my day in three short steps."
The cut
Thursday’s open drop was loud: Meta put 816B weights on the table, Nvidia answered with a rack-to-desk stack the same day, and StepFun added another big sparse MoE you can download. Mid-window, Grok Voice took the speech board at the same $0.08-a-minute price, and Anthropic told investors it booked $16.8B last quarter.
APIs you rent will keep getting better and more expensive to ignore. The weights you can keep — Muse Spark Open on iron you control, Bolt aimed at 24GB (32GB in practice), Step-5 on a multi-GPU box — are still the ones that do not vanish when the invoice changes.