Models / Meta
Llama 4 Maverick
by Meta ·
meta-llama/llama-4-maverickLlama 4 Maverick 17B Instruct (128E) is a high-capacity multimodal language model from Meta, built on a mixture-of-experts (MoE) architecture with 128 experts and 17 billion active parameters per forward...
tool-usevisionreleased 2025-04-05
Legal score · raw
-
not yet benchmarked
Context
524K
max output 16K
Input
$0.23
per 1M tokens
Output
$0.82
per 1M tokens
Suite cost
-
no run recorded
Prompt caching
Grounded answers re-send the same retrieved authority every turn, so a cached read is what a long matter actually costs. Prices are per 1M tokens, from the provider’s own catalog.
Cache read
$0.06
3.8x cheaper than uncached
Uncached input
$0.23
per 1M tokens
Capabilities
Modalities
in textimage
out text
Request parameters it accepts
frequency_penaltylogit_biaslogprobsmax_tokenspresence_penaltyrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_logprobstop_p
Benchmark results
This model has not been run through the suite yet. See Humanity's Last Lawsuit for the benchmark models are ranked on.
Measured by DocketBuster
These numbers come from DocketBuster's own legal battery, not from DocketRouter's suite. Latest run per metric, with n and a 95% Wilson interval where the source reports one. See docketbuster.com/benchmarks.
| Metric | Value | n | Interval | Measured |
|---|---|---|---|---|
| Statute pinpoint, exact section (no retrieval) | 19.0% (57/300) | 300 | 95% CI 15.0% to 23.8% | 2026-08-22 |
| Statute pinpoint, exact section (with DocketBuster retrieval) | 78.5% (241/307) | 307 | 95% CI 73.6% to 82.7% | 2026-08-24 |
| Say-nothing rate (declines to bluff when the answer is not in the record) | 100.0% | 218 | 95% CI 98.3% to 100.0% | 2026-08-22 |
| Abstained on statute pinpoint | 9.4% (29/307) | 307 | count, no interval reported | 2026-08-24 |
| Coaching quality (rubric, 0 to 8) | 6.56 / 8 | - | rubric mean, no interval reported | 2026-08-23 |