Open weights buy control. Decide which job needs it.

Two labs shipped open-weight coding models on back-to-back days last week: Meta's Muse Spark Open on Oct 15 and Reflection's Beam on Oct 16. Both have 1M-token context windows, weights on Hugging Face, and Terminal-Bench 2.1 scores of 76.4 and 80.1.

I think many builders will swap one in because they assume open means cheaper. The hosted price matches Meta's own closed tier. Open weights buy you control over where the model runs, what you change about it, and whether it disappears on someone else's schedule. That pays off only for the jobs in your product that need it.

The Signal

Muse Spark Open, from Meta Superintelligence Labs, is a mixture-of-experts model with 816B total parameters and 42B active, built for coding and agent work. It takes text, image, video, and audio in and returns text. Reasoning effort runs low, medium, high, and xhigh. The default is xhigh, so set it deliberately.

At xhigh, Meta's model card shows 76.4 on Terminal-Bench 2.1, 60 on DeepSWE 1.1, 89.2 on MCP-Atlas, and 80.4 on OSWorld Verified. The same table puts it at 39 on the Artificial Analysis Intelligence Index, behind GLM 5.3 at 45 and Kimi K3 at 44. Credit to Meta for printing that row. (Hugging Face)

Self-hosting needs at least 700GB of VRAM for the NVFP4 checkpoint and 2TB for BF16. Meta says it's tuned to run on two Mac Studios with M5 Ultra using MLX.

Hosted, it's $1.25 per million input tokens, $4.25 per million output, and $0.15 per million cached input through OpenRouter providers including Together, Fireworks, Baseten, and Modal. eesel AI notes that matches Meta's closed Muse Spark 1.3 Standard tier. Vercel AI Gateway carries it as meta/muse-spark-open, with zero data retention and no training on your data. (OpenRouter, Together AI, eesel AI, Vercel)

The Muse Model License lets you use, modify, fine-tune, distill, and sell. "Built with Muse Spark Open" must appear prominently in your product UI, docs, website, or settings, and that also applies to models trained or distilled from its outputs. If you sell API access to the model and your company group earns over $20M in 12 months, you have to notify Meta and sign a separate agreement. Internal use and features inside your own product don't trigger that clause. Its usage policy says it isn't intended for users under 18.

Beam is the first model from Reflection, a new US lab. It's 501B total parameters with about 23B active, text in and text out, 1M context, under Apache 2.0. It has five reasoning levels with medium as the default, plus a preserve_thinking option that carries reasoning across turns. It ships with a Rust-based harness and BF16, FP8, and NVFP4 checkpoints. Reflection reports 81.6% on SWE-Bench Verified and 80.1% on Terminal Bench v2.1. Meta's own table also lists Beam at 80.1, ahead of Muse. (Hugging Face, Java Rubber Duck)

Reflection's model card doesn't prohibit distillation, training other models on its outputs, fine-tuning, quantization, or merging. Ollama Cloud runs it free "for a limited time" with zero data retention, and no end date or paid price has been announced. Treat that as a test window and keep it out of your margin math. (Ollama, Ollama library)

The Stack

1. List every AI job in your app with the five fields the prompt below uses.
2. Mark the jobs that need control: data that can't leave your infrastructure, a fine-tune you plan to run, or a pinned model you can't lose to a deprecation.
3. Leave everything else on its current hosted closed model. If you can't name the control reason, the job stays put.
4. For any Muse candidate, decide where the "Built with Muse Spark Open" line goes before you build. Check the $20M clause only if you resell API access.
5. Next to what you pay now, write the hosted Muse price and the hardware self-hosting needs (700GB of VRAM for NVFP4, 2TB for BF16).
6. Before you buy hardware, run your current task set through meta/muse-spark-open on a gateway and beam:cloud on Ollama, and compare both with your current default. ollama launch claude --model beam:cloud runs Beam inside Claude Code, and the same launcher works with codex and opencode.

One decision: pick the single job where control matters most and move only that one. If no job has a control reason, move none.

Prompt of the Day

Paste this with the list of AI jobs in your app:

Here's a list of the AI jobs in my app. Each row has the job name, the data it touches, monthly volume, current model, and whether customers see the output.

Sort every job into one of two groups: "keep on hosted closed model" or "candidate for open weights." Give the reason for each: data sensitivity, a real need to fine-tune, or deprecation risk. If none of those apply, keep it hosted.

For every candidate, list the license obligations if I use Muse Spark Open. The line "Built with Muse Spark Open" must be displayed prominently in my product UI, docs, website, or settings, and that also applies to any model I train or distill from its outputs. Flag the $20M Model-as-a-Service clause only if I tell you I sell API access to the model. Internal use and features inside my own product don't trigger it.

Use only what I give you. Do not estimate volumes, costs, token counts, or savings. If a field is missing, write "missing."

If it puts a dollar figure on a job you never priced, throw the answer out.

Bring your sorted list to AI Freedom Lab, where builders launch companies and web and mobile apps with AI.

Tool of the Day

GitHub Copilot, featured in the AI Tools Index.

Copilot gets the slot because Gemini 4 Argon is now selectable inside it. On Copilot Business or Enterprise, Argon is on by default unless you disabled new models, and it bills at provider list price under usage-based billing. If you're a Business or Enterprise admin, open your model settings today and decide whether Argon stays on.

Quick Hits

OpenAI retired GPT-5.5 from Codex for ChatGPT sign-ins on Oct 14 at noon PDT. API and API-key Codex sessions aren't affected. Replace gpt-5.5 in workspace defaults, saved settings, managed configs, custom agents, scheduled tasks, and scripts with gpt-6.1-sol on Plus, Pro, Business, Enterprise, and Edu, or gpt-6-luna on Free and Go. Check scheduled jobs first, since a pinned model there fails without warning. (OpenAI changelog)

Claude Code 2.1.299 added idle compaction. A terminal conversation over 200K tokens that sits idle for about 54 minutes gets compacted just before its one-hour prompt cache expires, so you don't re-pay to cache the whole context when you return. Tune the threshold with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, or turn it off with "idleCompaction": false. (Claude Code changelog)

Gemini 4 Argon went GA in GitHub Copilot on Oct 13 for Pro+, Max, Business, and Enterprise. It's in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, and the coding agent. (GitHub)

Closing the loop

Any model you rent can be retired on someone else's clock. GPT-5.5 left Codex for ChatGPT sign-ins at noon on a Wednesday. Weights you host stay where you put them, and that's worth paying for on one job, maybe two, maybe none. Decide this week.

---

Join AI Freedom Lab, where builders launch companies and web and mobile apps with AI: skool.com/aifreedomlab