11 models · on the curve since 2023

Meta

Value position · the site’s read Mixed

Meta’s strength is three billion users and a strategy of commoditising rivals’ moats with open weights. The 2026 pivot to closed frontier models bets that distribution plus a private model beats giving the weights away.

What would change this readThe closed pivot fails to out-earn the open strategy, or a rival assistant displaces Meta’s apps as the place people spend attention.

Recent cadence
3 in last 12 mo
3 the year before
Landmark density
6 / 11
defining releases
Weights
Open + closed
Muse Spark 1.1 flagship

Founded Meta Platforms (2004)Menlo Park, USA

The substrate underneath figures as of Aug 2026

Trains and serves on

It owns the buildings and almost none of the silicon. A definitive multi-year agreement with AMD covers up to 6 GW of Instinct GPUs on the MI450 architecture, with the first gigawatt shipping in the second half of 2026 and AMD granting Meta a warrant over up to 160 million of its own shares against shipment milestones. Alongside that: a reported NVIDIA agreement, reported rental of Google TPUs, and its own MTIA accelerator for inference. Sites at New Albany, Ohio and Richland Parish, Louisiana, the latter expanded to 5 GW.

ConcentrationMeta owns its datacentres but almost none of its accelerators, leaving it simultaneously dependent on NVIDIA, on AMD, and on Google — the direct model competitor whose TPUs it is reported to rent.

Weights
Partial and narrowing, which is the story. The Muse frontier family — Spark, Image, Video — ships CLOSED. Llama remains open-weight under Meta's own community licence, which is not OSI open source and requires a separate agreement above 700M monthly users. The company that made open weights its position shipped its 2026 frontier model shut.
Where you buy it
meta.ai and the Meta AI app · Meta Model API (US developers, preview) · Llama via third-party clouds
Structure
Meta Platforms, Inc., a Delaware public company with dual-class stock. Class B carries ten votes a share, and Mark Zuckerberg holds roughly 61% of total voting power against about 14% of the economics.
Revenue
AI revenue is not disclosed separately; Meta reports only Family of Apps and Reality Labs. Group revenue was $60.8B in the quarter to 30 Jun 2026, up 28%, with net income below the prior year at $15.8B.
People
75,472 across Meta (30 Jun 2026), down 1% year on year; Superintelligence Labs is not broken out

Biggest single dependencyMeta is committing $130-145B of 2026 capex and has abandoned open weights at the frontier, while disclosing no AI revenue line and earning less than the prior year on 28% higher revenue.

No price is shown because none is confirmed. Muse Spark 1.1 is reported at $1.25/$4.25 per million tokens by several outlets, but Meta's own launch post states no prices and its pricing URL returns nothing, so the figure stays out of the headline position. Nor is there a trajectory to draw: two years ago Meta sold no model by the token at all — Llama was a free download — so the July 2026 API is a business-model entry, not a price cut. The often-quoted $600B three-year infrastructure figure is an executive statement; the filed number is the 2026 capex range above.

On the durability curve frontier era
2023202420252026 LLaMA 1Muse Spark 1.1 ARCHITECTURE the substrate · compounds  → OUTPUT per model · commoditises  → CROSSOVER the durability thesis  → DURABLE VALUE-CAPTURE ↑ the margin a release keeps · not raw capability TIME · FRONTIER MODEL RELEASES →
Output · value a model keeps Architecture · the compounding substrate Meta release

Hover any point to read the release. The y-axis is qualitative; only the x-axis carries dates.

When to reach for them practical read
Open at scale
Reach for Llama when you need frontier-scale open weights to self-host (Llama 3.1 405B, Jul 2024).
Fine-tuning base
Pick a Llama checkpoint as a base to adapt freely (Llama 2, first commercial open license, Jul 2023).
On-device
Use the smaller multimodal Llamas where the model must run locally (Llama 3.2, Sep 2024).
The landmark releases 6 of 11 models
  1. Jul 2026 Muse Image First image model out of Meta Superintelligence Labs ImageNew modality
  2. Apr 2026 Muse Spark First closed-weight Meta frontier model ClosedPivot
  3. Apr 2025 Llama 4 (Scout/Maverick) Mixture-of-experts, multimodal OpenMoEMultimodal
  4. Jul 2024 Llama 3.1 First open model at frontier scale Open405B128K ctx
  5. Jul 2023 Llama 2 First commercial open license OpenCommercial
  6. Feb 2023 LLaMA 1 Research weights that leaked and seeded the open scene OpenResearch
Full release logall 11, newest first
  1. 2026
  2. Jul Muse Spark 1.1 Multimodal reasoning for agentic work; 1M context; Meta Model API preview Agentic1M ctxCurrent
  3. Jul Muse Image First image model out of Meta Superintelligence Labs ImageNew modality
  4. Apr Muse Spark First closed-weight Meta frontier model ClosedPivot
  5. 2025
  6. Apr Llama 4 (Scout/Maverick) Mixture-of-experts, multimodal OpenMoEMultimodal
  7. 2024
  8. Dec Llama 3.3 70B efficiency update Open70B
  9. Sep Llama 3.2 First multimodal Llama OpenMultimodal
  10. Jul Llama 3.1 First open model at frontier scale Open405B128K ctx
  11. Apr Llama 3 8B and 70B open models Open8B / 70B
  12. 2023
  13. Aug Code Llama Code-specialised variant OpenCoding
  14. Jul Llama 2 First commercial open license OpenCommercial
  15. Feb LLaMA 1 Research weights that leaked and seeded the open scene OpenResearch
Where they sit on weights versus the field
The durability take on Meta from the archive
About Meta at a glance
Parent
Meta Platforms (2004)
HQ
Menlo Park, USA
Leadership
Mark Zuckerberg
Division
Superintelligence Labs
Flagship
Muse Spark 1.1
Weights
Open (Llama) → closed
Also builds
Facebook, Instagram, WhatsApp
Known for
Llama open-weight models

Meta Platforms (founded as Facebook, 2004, Menlo Park) builds AI through Meta Superintelligence Labs. It became known for the open-weight Llama family — the most widely adopted open LLMs — before pivoting in 2026 toward closed-weight frontier models.

The frontier chart → Read the archive →