K Koda Intelligence
Subscribe
Cover image for the 16 October 2026 Signal: Open weights, closed crown: Meta's 816-billion-parameter Muse Spark Open ships while its best models stay private.
THE SIGNAL · 16 OCT 2026 · 9 MIN READ

Meta releases Muse Spark Open weights and keeps its best models closed

Meta shipped Muse Spark Open, an 816-billion-parameter mixture-of-experts model with a 1-million-token context, while Bloomberg notes its strongest models stay private. PrismML's Bonsai 2 models go the other way, as small as 0.43GB.

KI Koda Editorial · Updated 01:20 GMT · 25 sources
AIOPEN WEIGHTSMARKETS
31 of 41 figures cited18 receipts in the register
  1. 816 billion1Bloomberg
  2. 42 billion2Bloomberg
  3. $993Aonic Club, via Superhuman
The register
Lead Story

The day's defining move.

Open Source · Bloomberg
Lead storyOpen Source 816 billion1 Bloomberg · 15 OCT 2026
Open Source·Bloomberg·15 OCT 2026

Meta Releases Muse Spark Open, Keeps Best Closed

Meta released Muse Spark Open weights on October 15, 2026: a mixture-of-experts model with 816 billion1 total and 42 billion2 active parameters and a 1-million-token context, built for coding agents. It ships via Hugging Face and Microsoft Foundry and runs locally on two M5 Ultra Mac Studios through Apple's MLX. Bloomberg notes Meta keeps its best models closed.

Why it mattersTeams planning to run Muse Spark Open should size hardware and per-token compute against its 42 billion2 active parameters, and should not expect it to match the closed models Meta keeps for itself.

Continue reading
Markets

Ether posted the session's largest move at +1.00%, with Bitcoin up 0.90%, WTI crude up 0.89%, the S&P 500 up 0.33% and the Nasdaq up 0.08%, while the crypto Fear & Greed index read 63 (Greed).

KODA PROFrom the desk

The Operator Tier is coming.

A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.

Join the founding list →
Listen & Watch

Daily broadcasts.

Podcast · Infographic
Visual

Intelligence map

Markets

6 levels at the close.

Market TerminalLAST CLOSE 15 OCT 2026 · READ 2026-10-16 · Yahoo Finance
S&P 500 7,744.0612 +0.33%12 6D, window short 7,718.94 low / 7,811.54 high
Nasdaq 26,713.0313 +0.08%13 6D, window short 26,692.87 low / 27,366.17 high
Bitcoin $82,294.2114 +0.90%14 6D, window short $81,562.49 low / $83,679.47 high
Ethereum $2,442.4415 +1.00%15 6D, window short $2,418.21 low / $2,534.70 high
Crude Oil (WTI) $95.9416 +0.89%16 6D, window short $91.49 low / $95.94 high
Crypto Fear & Greed 63 / 10017 Greed 0 to 100 index 25 fear / 50 neutral / 75 greed · alternative.me
Today's Focus

What else the day turns on.

01

Aonic Club's $993 Membership

Aonic Club sells its own branded products at the cost of manufacturing and shipping for a $993 annual membership. It claims prices run 20% to 80% below name-brand alternatives and pitches the service against Walmart, Amazon and Costco.

Source: archive.superhuman.ai
02

Anthropic's In-House Version Control

Anthropic built an in-house version control system after AI-generated code drove 30,0004 merged changes a day. At that volume, code written by AI became the reason to rebuild the tooling that tracks it.

Source: anthropic.com
Deep Dive

Today's editorial.

All editorials
Deep Dive cover image: Meta Gave Away Its Second-Best Model
Deep Dive7 min read

Meta Gave Away Its Second-Best Model

Read full analysis
The Wire

3 stories from 3 desks.

Lead carried above
Open Source·PR Newswire·15 OCT 2026

PrismML Shrinks Qwen3.5 Into Ternary Bonsai 2 Models

At 0.43GB, 0.95GB and 2.00GB, PrismML's new Bonsai 2 2B5, 4B6 and 9B7 ternary models, based on Qwen3.5, are over 9x8 smaller than full precision and ship under Apache 2.0.

Why it mattersPhone and smartglasses developers should run coding and math on Bonsai 2 locally, where PrismML claims parity, and keep tool calling on full-precision Qwen3.5.

Read
Enterprise·CNBC·13 OCT 2026

JPMorgan Blocks Outside AI Agents From Customer Accounts

JPMorgan is blocking outside AI agents from directly accessing customer accounts. The same week, Shopify widened its catalog to personal agents while keeping human sign-off.

Why it mattersConsumer-finance agents that log into JPMorgan accounts on a customer's behalf should be treated as broken, while shopping agents can still reach Shopify's catalog as long as a human signs off on.

Read
Enterprise·OpenAI

OpenAI Launches Rosalind Academic Plan for Universities

OpenAI launched the Rosalind Academic Plan for universities. It offers resources and support to academic institutions running AI research.

Why it mattersHold campus AI-tool renewals until OpenAI publishes Rosalind's pricing and terms, since the plan could undercut licenses universities already pay for.

Read
The Globe

2 stories off the AI beat.

Conflict·Reuters·15 OCT 2026

Ukraine Hits Russian Refinery as Energy Truce Frays

Kyiv had signalled support for the U.S.-announced pause on energy strikes. After Russia's large-scale strike on its power grid, Ukraine hit a Russian refinery and said it will not honor the pause unless Moscow does.

Why it mattersGrids and refineries on both sides are back in range, so fuel and power hedges priced on the truce holding need repricing before winter demand peaks.

Read
Also on the globe
  1. ReutersJapan Urges G20 to Watch Currency Moves Urgently
The Lab

Tool of the day, researched.

All Lab reports
Eido home page, captured for the 16 October 2026 Lab report Tool of the day 6.418/ 10
Coding Free (beta)

Eido

Eido puts the Pi coding agent inside a Zed-based Mac editor where its edits land in a review view, and every line of the app is readable on GitHub.

Eido's pitch is legibility: the agent works inside a native editor, its file changes surface for inspection, and the source sits in a public repository you can audit. It is a one-maintainer beta with limited language support, so treat it as an afternoon experiment for Mac developers curious about Pi and keep it away from production work for now.

Capability 6.0
Ease 5.0
Value 8.0
Momentum 6.5
Read the Lab report Try it
Also on the radar
Productivity

The Agentic Sales Platform: outbound sales signals, data, and automation in one place

This platform combines real-time buying signals, verified contact data, waterfall enrichment, LinkedIn automation, email automation, and deliverability infrastructure, so outbound teams can stop stitching together separate tools. Start by mapping your current prospecting stack against its feature list and pilot one campaign end to end to see which subscriptions it could replace. Pricing is by quote, so bring your seat count and monthly send volume to the sales call.

Try it
Productivity

Usermaven: AI marketing attribution that ties channels to revenue

Usermaven gives B2B SaaS companies, agencies, and ecommerce brands multi-touch attribution, customer journey analytics, and AI-generated insights about which channels actually drive revenue. Connect your site and conversion events, then compare its multi-touch view against your current last-click reports to find channels you are undervaluing. It holds a 4.8 rating from 19 Product Hunt reviews, and pricing is not listed, so confirm plan costs before you commit.

Try it
Creativity

AI Soundtrack: generate original background music fitted to your video

AI Soundtrack v2.0.0 lets you upload a video, choose a mood, and get an original instrumental track timed to the length of your clip. Use it for social cuts, product demos, and tutorials where stock music either clashes with the pacing or brings licensing questions. Generate two or three moods for the same clip and pick the one that matches your edit points best.

Try it
Productivity

Superhuman AI: AI features built to get you to inbox zero faster

Superhuman AI is a suite of AI features inside the Superhuman email client, designed to help you save time and clear your inbox faster. Try it on your heaviest email day of the week and track how many threads you close compared with a normal day. If the time savings hold up over a week, build the AI features into your daily triage routine.

Try it
The Arena

2 labs moved this week.

  1. Meta AIMeta released Muse Spark Open, an 816-billion-parameter open-weight multimodal model with 42 billion2 active parameters and a 1-million-token context window. straitstimes.com · 15 OCT 202601
  2. OpenAIOpenAI launched the Rosalind Academic Plan to give academic institutions resources and support for AI-driven research. openai.com02
Builder Radar

1 receipt from the repos, one paper.

Primary sources, pulled by index
Receipts
  1. Behind: Meta Releases Muse Spark Open, Keeps Best Closed DocsWelcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

    Muse Spark Open is a coding and agentic model. We wanted to see that for ourselves, so we tried a mixture of multi-modal and agentic tasks that use the model's image and audio inputs. Everything below is something you can run or fork.

    github.com
Paper of the day

Apertus 1.5: From Open Text Models to Open Multimodal Reasoners

Project Apertus, :, Yixuan Xu et al. · arXiv 2026-10-11

Language models are becoming infrastructure, and infrastructure built in the open, by public institutions, on documented data, is the alternative to dependence on a few providers; Apertus shows this is viable at scale.

The 3,600-token filter costs little separation: the mean per-prompt max-min gap is 3.822 before filtering against a realized 3.766 after it.
Read the paper

Receipts come from the Firecrawl Developer Index (READMEs, docs, issues, merged PRs), first-party sources only. The paper comes from the Firecrawl Research Index: arXiv, indexed in the last 7 days and submitted this month or last, picked for today's lead. Quotes are verbatim.

THE DOJO · BUILD TODAY

Test open weights at two sizes before you commit

  1. 01

    Rent a trial of Muse Spark Open. Pull it through Hugging Face or Microsoft Foundry and run one real coding-agent task from your backlog. Record quality and latency before you price any hardware.

  2. 02

    Run a ternary Bonsai 2 model locally. Download the 4B version at 0.95GB, which ships under Apache 2.0, and point it at one routine task in your pipeline. If it holds up, you have a cheap local fallback.

  3. 03

    Review an agent's diffs in Eido. Open a throwaway repository, let the agent make changes, and read each surfaced file change before accepting it. Check that your language is supported, since this is a one-maintainer beta.

The Bottom Line

Open weights now come in two sizes: rent the giant, run the small.

Meta's Muse Spark Open shows that a big lab can release a headline model while its best work stays closed, and PrismML's Bonsai 2 shows how small a capable open model can get. Together they suggest matching the model to the job, not chasing the largest checkpoint. Trial the 816-billion-parameter model on rented capacity for hard agent work, and keep a sub-1GB ternary model on hand for routine local tasks. Revisit the split once you have measured quality on your own workload.

Receipts

Every number, and where it came from.

31 of 41 figures cited, 18 receipts: 0 verified, 16 reported, 2 estimated
  1. 01816 billion...on October 15, 2026: a mixture-of-experts model with 816 billion total and 42 billion active parameters and a...Bloomberg15 OCT 2026as publishedR
  2. 0242 billion...a mixture-of-experts model with 816 billion total and 42 billion active parameters and a 1-million-token context...Bloomberg15 OCT 2026as publishedR
  3. 03$99Aonic Club launches a $99 subscription to sell goods at manufacturing and shipping costAonic Club, via Superhuman15 OCT 2026as publishedR
  4. 0430,000Anthropic built an in-house version control system after AI-generated code drove 30,000 merged changes dailyAnthropic, via TLDR15 OCT 2026as publishedR
  5. 052BAt 0.43GB, 0.95GB and 2.00GB, PrismML's new Bonsai 2 2B, 4B and 9B ternary models, based on Qwen3.5, are over 9x...PR Newswire15 OCT 2026as publishedR
  6. 064BAt 0.43GB, 0.95GB and 2.00GB, PrismML's new Bonsai 2 2B, 4B and 9B ternary models, based on Qwen3.5, are over 9x...PR Newswire15 OCT 2026as publishedR
  7. 079B...0.43GB, 0.95GB and 2.00GB, PrismML's new Bonsai 2 2B, 4B and 9B ternary models, based on Qwen3.5, are over 9x smaller...PR Newswire15 OCT 2026as publishedR
  8. 089x...2 2B, 4B and 9B ternary models, based on Qwen3.5, are over 9x smaller than full precision and ship under Apache 2.0.PR Newswire15 OCT 2026as publishedE
  9. 09$1BBumble in advanced talks to go private at $1B valuationTLDR15 OCT 2026as publishedR
  10. 10$500MAI Agent Developer Manus Raises $500M+ at Reported $4B ValuationTLDR15 OCT 2026as publishedR
  11. 11$4BAI Agent Developer Manus Raises $500M+ at Reported $4B ValuationTLDR15 OCT 2026as publishedR
  12. 127,744.06S&P 500 close, +0.33% on the sessionYahoo Finance15 OCT 2026yfinance closeR
  13. 1326,713.03Nasdaq close, +0.08% on the sessionYahoo Finance15 OCT 2026yfinance closeR
  14. 14$82,294.21Bitcoin close, +0.90% on the sessionYahoo Finance16 OCT 2026yfinance closeR
  15. 15$2,442.44Ethereum close, +1.00% on the sessionYahoo Finance16 OCT 2026yfinance closeR
  16. 16$95.94Crude Oil (WTI) close, +0.89% on the sessionYahoo Finance15 OCT 2026yfinance closeR
  17. 1763 / 100Crypto Fear & Greed close, Greed on the sessionalternative.me16 OCT 2026as publishedR
  18. 186.4 / 10Eido, Koda score across four dimensions04R dossier16 OCT 202604R dossier scoringE
V
Verified: the stat gate corroborated this figure against an independent search before publication.
R
Reported: the figure is carried as published by the linked source and was not independently corroborated.
E
Estimated: the figure is approximate or hedged, either in the source or by the stat gate.

Method names the computation path behind the row, from a closed vocabulary: as published, yfinance close, 03B search corroboration, 04R dossier scoring, not recorded.

A figure is money, a percentage, a spelled magnitude, a multiple, a rate carrying its unit, a score over its denominator, a thousands separated number or an index close. Bare integers and years are not counted, so a phrase like "over 100 integrations" carries no mark. A chart's own scale line is the axis of the level above it and is counted once, with that level.

Get the morning Signal

203 editions so far, one a day. Unsubscribe anytime.

Forward this to one operator you work with. Your referral link is in every email; milestones at koda.community/refer.