Meta

Key Terms

• Open-weight model - model weights are published for download, allowing anyone to run inference, fine-tune, or modify the model locally or on their own infrastructure. The Llama 4 Community License imposes restrictions on derivative naming and has a 700M monthly active user commercial threshold. Source: GitHub – License
• Mixture-of-experts (MoE) - architecture where only a subset of parameters are activated per token. Llama 4 always activates 17B parameters per forward pass but routes through different expert subsets (16 for Scout, 128 for Maverick). This keeps inference cost low despite large total parameter counts. Source: Openrouter – Llama 4 Maverick
• Early fusion - a multimodal training technique where text and vision data are processed together from the first layer, rather than using separate frozen vision encoders. Llama 4 uses early fusion for native multimodality.
• Llama 4 Community License - Meta's custom license for Llama 4 models. Permits free commercial use with two conditions: (1) entities with more than 700M monthly active users must request a separate license from Meta, and (2) derivative models must prefix their name with "Llama" and display "Built with Llama" on related materials. Source: GitHub – License
• Llama API - Meta's hosted inference API at llama.developer.meta.com, currently in waitlist mode. Not yet publicly available. Source: Llama
• Context window - the maximum number of tokens a model can process in one conversation. Maverick supports 1M tokens; Scout supports 10M tokens. Source: Openrouter – Llama 4 Scout

Latest Changes

Changes since the 2026-05 report.

• No new models. Llama 4 Scout and Llama 4 Maverick remain the flagship models, released April 5, 2025. No Llama 4.1, Llama 5, or Llama 4 Behemoth was announced or released in June 2026. The models are now 14+ months old. Source: Openrouter – Llama 4 Maverick
• No first-party pricing changes. Meta does not sell API access, so there are no pricing changes from Meta. The Llama API remains in waitlist mode with no public pricing or timeline. Source: Llama
• Price change (third-party): Scout input up on OpenRouter. OpenRouter's listed Scout input price rose from $0.08 per MTok (May report) to $0.10 per MTok as of June 30, 2026, a 25% increase. Output is unchanged at $0.30 per MTok. Source: Openrouter – Llama 4 Scout
• Price unchanged (third-party): Maverick stable. Maverick remains $0.15 input / $0.60 output per MTok on OpenRouter, unchanged from May. Source: Openrouter – Llama 4 Maverick
• Llama 4 still not on Together AI. As of June 30, 2026, Together AI's pricing page lists Llama 3.3 70B ($1.04 / $1.04 per MTok) and Llama 3 8B Instruct Lite ($0.14 / $0.14 per MTok) but does not carry Llama 4 Scout or Maverick. Source: Together – Pricing
• Llama API waitlist persists. No update on pricing, availability, or capabilities of the first-party Llama API. The waitlist has been open since at least April 2025. Source: News – Item
• No Llama-specific blog posts in June 2026. Meta's AI blog (ai.meta.com/blog/) could not be fetched this cycle (see Failed Sources), but no Llama model launch or feature announcement was surfaced via HN, OpenRouter, or community channels during June.
• Behemoth still MIA. Llama 4 Behemoth, mentioned at the April 2025 launch as a larger model in training, has had no status update in over 14 months. Source: Llama
• Benchmark credibility flagged (January, still unrebutted). A January 2026 report quoted departing Meta AI leadership confirming Llama 4 benchmark numbers were "fudged a little bit." Meta has not issued a public correction or restated scores. Source: Tech – Results Were Fudged Departing Meta Ai Chief Confirms Llama 4 Benchmark Manipulation

Plans

Meta does not offer subscription plans, tiers, or bundled products. Models are free to download and use under the Llama 4 Community License.

API Pricing

Meta does not publish API pricing because it does not offer a public API. Third-party providers set their own prices for serving Llama 4 models. All prices below are per 1M tokens.

Self-hosting hardware requirements:

• Llama 4 Scout: Can run on a single H100 GPU with INT4 quantization (109B total params). Source: Openrouter – Llama 4 Scout
• Llama 4 Maverick: Cannot run on a single GPU. FP8 quantized weights fit on a single H100 DGX host (8 GPUs). BF16 weights require multi-host deployment (400B total params). Source: Hugging Face – Llama 4 Maverick 17B 128E

Model Performance / Benchmarks

Meta reports the following benchmarks for instruction-tuned Llama 4 models. These scores are taken from Meta's published materials. Note: departing Meta AI leadership confirmed in January 2026 that some Llama 4 benchmark numbers were "fudged a little bit," so these figures should be treated with caution. Source: Tech – Results Were Fudged Departing Meta Ai Chief Confirms Llama 4 Benchmark Manipulation

Context for comparison: Llama 4 Maverick's LiveCodeBench score of 43.4 (now under a benchmark-manipulation cloud) is well below current closed-source frontier models. At $0.15 / $0.60 per MTok via OpenRouter, Maverick still offers competitive performance-per-dollar for cost-sensitive workloads, but is no longer a frontier-class coding model. Source: Openrouter – Llama 4 Maverick

Latest News

No Llama Successor Launched (June 2026)

June 2026 passed with no Llama 4.1, Llama 5, or Llama 4 Behemoth announcement. HN searches for "Llama 4" and "Meta Llama" surfaced no new launch threads dated June 2026. The two current models remain Scout and Maverick, both released April 5, 2025, making them 14+ months old. Source: Hn – Search

Llama API Still in Waitlist Mode (ongoing)

The first-party Llama API at llama.developer.meta.com remains waitlist-only with no public pricing, rate limits, or general-availability timeline. Community reports confirm the waitlist has persisted for over 14 months since Llama 4 shipped. Source: News – Item

Benchmark Manipulation Confirmed by Departing Meta AI Leadership (January 2026, unrebutted)

In January 2026, departing Meta AI leadership confirmed that Llama 4 benchmark numbers were "fudged a little bit," according to an FT report syndicated by Slashdot. Meta has not issued a public correction, restated its scores, or published corrected benchmark results in the six months since. This is relevant because the published LiveCodeBench, MMLU Pro, and GPQA Diamond numbers above have not been independently verified or re-attested by Meta. Source: Tech – Results Were Fudged Departing Meta Ai Chief Confirms Llama 4 Benchmark Manipulation

Llama 4 Behemoth Status Unknown (ongoing)

Llama 4 Behemoth, referenced at the April 2025 launch as a larger model in training, has had no status update, capability disclosure, or release timeline in over 14 months. Source: Llama

Community Signals

No significant Llama 4 community discussion in June 2026. HN searches for "Llama 4" (711 total stories) and "Meta Llama" (346 total stories) surfaced no new high-engagement threads dated June 2026. Community attention has shifted to the Sonnet 5 launch, DeepSeek V4 Pro pricing, GLM-5.2, and the GitHub Copilot AI Credits transition. Llama 4 has settled into a stable but low-attention "commodity" position in the open-weight ecosystem.

The most relevant recent community discussions (both from earlier in 2026, reflecting the dominant sentiment that persists into June):

"Ask HN: 10 months since the Llama-4 release: what happened to Meta AI?" (55 points, 12 comments, February 2026) captured the prevailing frustration with Meta's post-Llama-4 silence. The submitter noted the API was "still waitlist-only 10 months on." Key comments:

• hasperdi: "AFAIK Zuck got mad and restructured the whole department. What's likely is that there won't be anything open / significant coming out from them anymore." News – Item
• verdverm: "those people they hired at astronomical figures are just fucking around because they know Marc will look bad if he fires them. The most likely reason is internal dysfunction, they certainly have the resources to keep the same release pace." News – Item
• casey2: "Weird that Zuck is still huffing the open platform copium when everyone outside the US uses deepseek." News – Item
• BoredPositron: "History suggests that when Zuck takes a personal interest in a project, it tends to derail. Metas AI initiative might be headed for the same fate as his previous obsessions." News – Item

Source: News – Item

Benchmark manipulation confirmation erodes trust (30 points, January 2026). After the FT/Slashdot report that departing Meta AI leadership confirmed Llama 4 benchmarks were "fudged a little bit," the HN thread was short but pointed:

• DivingForGold: "Can you really trust ANYTHING from Meta?" News – Item

Source: News – Item

opencode harness pattern for open-weight models remains the primary way developers use Llama 4 for coding. The May 2026 DeepClaude/DeepSeek HN thread validated the opencode harness for open-weight and cheap-API models, and the same pattern applies to Llama 4 via OpenRouter or self-hosting. Developers route to Llama 4 when they need data sovereignty or want to avoid vendor lock-in, but few reported it as a primary coding model in June. Source: News – Item

Enterprise Readiness

Source: GitHub – License ; Openrouter – Llama 4 Maverick

Transparency Gaps