Llama 4 Maverick — reviews, specs & pricing
Meta's natively multimodal MoE with 400B total / 17B active parameters, Llama 4 Community License.
Summary
Llama 4 Maverick is a state-of-the-art natively multimodal Mixture-of-Experts (MoE) model developed by Meta, engineered to excel in high-performance chat and code generation tasks. By utilizing a massive 400-billion total parameter architecture that only activates 17 billion parameters per token, it delivers top-tier intelligence with remarkable computational efficiency. Released under the Llama 4 Community License, this model bridges the gap between massive-scale reasoning and accessible, cost-effective deployment.
Sample use case
A technology company integrates Llama 4 Maverick into their collaborative software development platform to power an intelligent, real-time programming assistant. Software engineers leverage its combined chat and coding capabilities to generate complex multi-file codebases, explain legacy algorithms, and troubleshoot bugs using interactive dialogue. Because the model only activates 17B parameters during inference, the company can host it locally on standard enterprise hardware to protect proprietary intellectual property while keeping API latency minimal for hundreds of concurrent users.
Specifications
- Provider: meta
- License: source-available
- Parameters: 400B
- Context: 1000k tokens
- Released: 2025-04-05
Pros
- Highly efficient MoE architecture with 17B active parameters
- Natively optimized for advanced chat and programming workflows
- Flexible deployment under the Llama 4 Community License
- Massive 400B parameter capacity for deep reasoning
Cons
- High storage footprint required for the full 400B weights
- License limitations may apply for massive commercial deployments
- Complex hosting infrastructure needed for optimal MoE routing
Average rating 2.2 from 5 community reviews on Reviuws.
Community reviews
Simon Willison's blog on Llama 4 Maverick
Rating: 3.0 / 5 — by Simon Willison (use case: long-context multimodal tasks)
Willison criticises Meta for dropping such a significant release on a weekend, but lays out Maverick's specs — 400B total, 128 experts, 17B active, 1M token context, multimodal — as genuinely notable pending hands-on testing.
Pros: Million-token context with text and image input; efficient mixture-of-experts design.
Cons: Weekend release hindered community evaluation; coverage relies on Meta's own post.
Ars Technica on Llama 4 Maverick
Rating: 2.0 / 5 — by Benj Edwards (use case: general benchmarking)
Ars reports the surprise release exposed a gap between Meta's ambition and reality, with the touted 10-million-token context proving elusive in practice and early independent tests leaving experts disappointed.
Pros: Advertises industry-leading context and MoE efficiency; multimodal out of the box.
Cons: Early tests disappointed relative to claims; the huge context proved hard to realise.
The Verge on Llama 4 Maverick
Rating: 2.0 / 5 — by Kylie Robison (use case: benchmark integrity)
The Verge reports Meta submitted a specially tuned, non-public experimental variant to LMArena that differed from the released weights, concluding Meta made Maverick look more competitive than the shipped model.
Pros: Public release still ships open weights with multimodal support.
Cons: Benchmarked variant differed from released weights; undermines trust in its leaderboard rank.
The Register on Llama 4 Maverick
Rating: 2.0 / 5 — by The Register staff (use case: release transparency)
The Register describes the LMArena episode as a bait-and-switch in which Meta uploaded a non-public, differently tuned variant to boost its rank, casting doubt on the model's benchmark standing.
Pros: Still an open-weight multimodal mixture-of-experts release.
Cons: Leaderboard variant gave longer, emoji-heavy answers unlike the shipped model; credibility damaged.
VentureBeat on Llama 4 Maverick
Rating: 2.0 / 5 — by Carl Franzen (use case: early production deployment)
VentureBeat reports Meta defending the release after reports of mixed quality, blaming inconsistent output on implementation bugs at inference providers rather than the models themselves.
Pros: Meta acknowledged and responded to feedback; some inconsistency may be provider bugs.
Cons: Early users reported mixed-quality output; unclear how much is bugs versus model limits.
Frequently asked questions
What is Llama 4 Maverick?
Llama 4 Maverick is a state-of-the-art natively multimodal Mixture-of-Experts (MoE) model developed by Meta, engineered to excel in high-performance chat and code generation tasks. By utilizing a massive 400-billion total parameter architecture that only activates 17 billion parameters per token, it delivers top-tier intelligence with remarkable computational efficiency. Released under the Llama 4 Community License, this model bridges the gap between massive-scale reasoning and accessible, cost-effective deployment.
How much does Llama 4 Maverick cost?
Pricing for Llama 4 Maverick is not publicly listed.
Is Llama 4 Maverick open source?
Llama 4 Maverick is released under the source-available license.