MDL
Llama 3.2 Instruct 90B (Vision)
Meta
Verified against Artificial Analysis · Jul 13, 2026
Capability
52
percentile index
Price /M
$1.38 / $1.38
in / out
Context
—
Avg score
37.3
6 benchmarks
AA Intelligence
6.2
index
Speed
61 tok/s
TTFT 0.6s
Released
Sep 25, 2024
About
Llama 3.2 Instruct 90B (Vision) is a open-weights model in the Llama family from Meta. Benchmarked on 6 evals, averaging 37.3.
Benchmark scores · 6
avg 37.3 — higher is betterThe research behind this model
all papers mentioning it →Information-Aware KV Cache Compression for Long Reasoning▲ 11 on HF · Jun 25, 2026Off-the-Shelf LLMs as Process Scorers: Training-Free Alternative to PRMs for Mathematical Reasoning▲ 7 on HF · Jun 1, 2026BioRefusalAudit: Auditing Biosecurity Refusal Depth Using General and Domain-Fine-Tuned Sparse AutoencodersMay 28, 2026Analyzing Quality-Latency-Resource Trade-offs in a Technical Documentation RAG Assistant Using LoRA AdaptationMay 27, 2026NestedKV: Nested Memory Routing for Long-Context KV Cache CompressionMay 26, 2026
Mentions matched by name in title/abstract — from the arXiv + HF daily corpus.
Performance & price
Output speed61 tok/s
Latency (TTFT)0.6s
Price in / out$1.38 / $1.38 per 1M
Blended price$1.38 / 1M
Median across API providers, via Artificial Analysis. Blended = 3:1 input:output. For reasoning models, latency includes thinking time. Methodology →
History
full ledger →Capability over time
Append-only ledger — every observed change to this model's numbers.
API providers
Artificial Analysis$1.38/M