GPT-5.6 (Sol / Terra / Luna) is now evaluated on TrustVector โ€” with day-1 independent verification, incl. METR's benchmark-cheating findings.

Read the evaluation
Evaluation record ยท llama-4-behemoth

Llama 4 Behemoth

v2025-02

Meta

Modelunreleasedopen-sourceself-hostedmathematics
85
Strong
About This Model

Meta's announced 2T-total/288B-active parameter Llama 4 teacher model that was NEVER RELEASED. It remains 'announced, not released' as of July 2026 โ€” effectively shelved (never formally cancelled): Meta gave no update when asked in January 2026 and has effectively exited open-weight frontier releases, shipping the proprietary closed-weight 'Muse Spark' (April 8, 2026) instead. Scores reflect unverifiable preview-era claims; the model is not available for any deployment.

Last Evaluated: July 9, 2026
Official Website

Trust Vector Analysis

Dimension Breakdown

๐Ÿš€Performance & Reliability
+

Preview-era claims: exceptional mathematical reasoning (95% MATH) and strong general knowledge (73.7% MMLU). The model was never released, so these results cannot be independently verified.

task accuracy code

Industry-standard coding benchmarks

Evidence
HumanEval Benchmark โ€” 75% pass rate (estimated from MATH performance)
MBPP Benchmark โ€” 82% on programming problems
highVerified: 2026-07-09
task accuracy reasoning

Advanced mathematical and scientific reasoning benchmarks

Evidence
MATH Benchmark โ€” 95% on mathematical reasoning tasks (industry leading)
GPQA Diamond โ€” 78% on PhD-level science questions
highVerified: 2026-07-09
task accuracy general

Crowdsourced comparisons and knowledge testing

Evidence
MMLU Benchmark โ€” 73.7% on multitask language understanding
LMSYS Chatbot Arena โ€” 1310 ELO (Top 5 overall)
highVerified: 2026-07-09
output consistency

Internal testing with repeated prompts

Evidence
Meta Internal Testing โ€” High consistency across diverse prompts
mediumVerified: 2026-07-09
latency p50

Median latency on recommended hardware

Evidence
Community benchmarking โ€” ~2.8s on standard hardware (self-hosted)
mediumVerified: 2026-07-09
latency p95

95th percentile response time

Evidence
Community benchmarking โ€” p95 latency ~5.2s (hardware dependent)
mediumVerified: 2026-07-09
context window

Official specification from provider

Evidence
Meta Documentation โ€” 128K token context window
highVerified: 2026-07-09
uptime

User-controlled deployment

Evidence
Self-hosted model โ€” Uptime depends on hosting infrastructure
Wikipedia - Llama (language model) โ€” Model was never released; weights are not available for any deployment as of June 2026
highVerified: 2026-07-09
๐Ÿ›ก๏ธSecurity
+

Good baseline security with self-hosted deployment offering full control. Additional safety layers recommended for production.

prompt injection resistance

Testing against prompt injection attacks

Evidence
Meta Safety Testing โ€” Good resistance, requires additional safeguards in deployment
mediumVerified: 2026-07-09
jailbreak resistance

Testing against adversarial prompts

Evidence
Meta Safety Evaluations โ€” Built-in safety mechanisms, additional layers recommended
mediumVerified: 2026-07-09
data leakage prevention

Analysis of deployment model

Evidence
Self-hosted deployment โ€” Full control over data in self-hosted deployments
highVerified: 2026-07-09
output safety

Safety testing across harmful content categories

Evidence
Meta Safety Benchmarks โ€” Safety training applied, additional filtering recommended
mediumVerified: 2026-07-09
api security

Review of deployment best practices

Evidence
Deployment documentation โ€” Security depends on deployment implementation
highVerified: 2026-07-09
๐Ÿ”’Privacy & Compliance
+

Exceptional privacy with self-hosted deployment. Full control over data residency, retention, and compliance. No data shared with Meta.

data residency

Analysis of deployment model

Evidence
Open-source model โ€” Full control over data location in self-hosted deployments
highVerified: 2026-07-09
training data optout

Analysis of data flow

Evidence
Self-hosted model โ€” No data sent to Meta in self-hosted deployments
highVerified: 2026-07-09
data retention

Analysis of deployment model

Evidence
Self-hosted deployment โ€” Full control over data retention policies
highVerified: 2026-07-09
pii handling

Review of deployment architecture

Evidence
Self-hosted deployment โ€” PII handling fully controlled by deployment team
highVerified: 2026-07-09
compliance certifications

Review of deployment options

Evidence
Self-hosted model โ€” Compliance achieved through deployment infrastructure
highVerified: 2026-07-09
zero data retention

Analysis of deployment model

Evidence
Self-hosted deployment โ€” Complete control over data retention
highVerified: 2026-07-09
๐Ÿ‘๏ธTrust & Transparency
+

Strong transparency as open-source model. Good training data disclosure. Customizable guardrails for specific use cases.

explainability

Evaluation of reasoning transparency

Evidence
Model Behavior โ€” Good explanations, strong mathematical reasoning transparency
mediumVerified: 2026-07-09
hallucination rate

Community evaluation and testing

Evidence
Community Testing โ€” Good factual accuracy, especially in mathematics
mediumVerified: 2026-07-09
bias fairness

Evaluation on bias benchmarks

Evidence
Meta Responsible AI Report โ€” Bias testing and mitigation applied
mediumVerified: 2026-07-09
uncertainty quantification

Qualitative assessment

Evidence
Model Behavior โ€” Good uncertainty expression
mediumVerified: 2026-07-09
model card quality

Review of documentation

Evidence
Meta Model Card โ€” Comprehensive model card with detailed benchmarks
highVerified: 2026-07-09
training data transparency

Review of technical documentation

Evidence
Meta Technical Report โ€” Good transparency on training methodology and data sources
highVerified: 2026-07-09
guardrails

Review of open-source safety systems

Evidence
Open-source implementation โ€” Transparent, customizable safety mechanisms
highVerified: 2026-07-09
โš™๏ธOperational Excellence
+

Operational scores are largely theoretical: the model was never released, so no deployment, support, or ecosystem exists for it. Meta shipped the closed-weight Muse Spark (April 2026) instead.

api design quality

Review of API design

Evidence
Meta Documentation โ€” Standard inference API, OpenAI-compatible
highVerified: 2026-07-09
sdk quality

Review of official and community SDKs

Evidence
Meta GitHub โ€” Official libraries and extensive community tools
highVerified: 2026-07-09
versioning policy

Review of versioning approach

Evidence
Meta Release Policy โ€” Clear model versioning and release notes
highVerified: 2026-07-09
monitoring observability

Review of available monitoring tools

Evidence
Community tools โ€” Observability depends on deployment stack
mediumVerified: 2026-07-09
support quality

Assessment of support channels

Evidence
Community Support โ€” Active community, official documentation
SiliconANGLE โ€” Release postponed in 2025; Meta provided no update when asked in January 2026
highVerified: 2026-07-09
ecosystem maturity

Analysis of ecosystem

Evidence
Open-source ecosystem โ€” Mature ecosystem with extensive tooling
Wikipedia - Llama (language model) โ€” No ecosystem exists for Behemoth itself; the model was never released and Meta has pivoted to closed-weight models (Muse Spark, April 2026)
AI CERTs News - Meta Behemoth Cancel Claim โ€” Re-verified July 2026: Behemoth remains unreleased and effectively shelved (never formally cancelled); reported root causes were mid-training MoE-routing changes and chunked-attention issues at 2T scale
highVerified: 2026-07-09
license terms

Review of license terms

Evidence
Meta Llama License โ€” Permissive commercial license
highVerified: 2026-07-09
Strengths
  • +Industry-leading mathematical reasoning (95% MATH)
  • +Strong general knowledge (73.7% MMLU)
  • +Complete data sovereignty with self-hosted deployment
  • +Open-source model with full transparency
  • +No data retention or sharing concerns
  • +Can achieve HIPAA and other compliance requirements
Limitations
  • !Requires significant infrastructure for deployment
  • !Higher latency than smaller models (~2.8s p50)
  • !Uptime and performance depend on hosting infrastructure
  • !Requires expertise to deploy and maintain
  • !No managed API service from Meta
  • !Large model size requires substantial compute resources
  • !Never released: still announced-only as of July 2026; Meta gave no update in January 2026 and pivoted to the closed-weight Muse Spark (April 8, 2026), so weights are unavailable
Metadata
pricing
input: Self-hosted (infrastructure costs)
output: Self-hosted (infrastructure costs)
notes: Open-source model, costs based on hosting infrastructure. Typically $0.50-2.00 per 1M tokens with optimized deployment.
context window: 128000
languages
0: English
1: Spanish
2: French
3: German
4: Italian
5: Portuguese
6: Japanese
7: Korean
8: Chinese
9: Arabic
10: Hindi
11: Russian
12: 100+ languages
modalities
0: text
api endpoint: Self-hosted
open source: true
architecture: Transformer-based, optimized for reasoning
parameters: 405B (estimated)

Use Case Ratings

code generation

Strong coding capabilities. Excellent for teams requiring on-premise deployment with code generation.

customer support

Good for customer support with self-hosted deployment for data privacy.

content creation

Strong content creation with excellent knowledge base (73.7% MMLU).

data analysis

Exceptional mathematical reasoning (95% MATH) ideal for complex data analysis.

research assistant

Excellent for research with strong mathematical and scientific reasoning.

legal compliance

Strong choice for legal applications requiring on-premise deployment and data sovereignty.

healthcare

Excellent for healthcare with self-hosted deployment enabling HIPAA compliance.

financial analysis

Outstanding mathematical reasoning (95% MATH) ideal for financial modeling.

education

Excellent for education, especially STEM subjects. Strong mathematical reasoning.

creative writing

Good creative writing capabilities, though not the primary strength.