004
Models
— Every model, one engine, one bill.
9 models ready to call, ours and the best open weights, each priced per million tokens and measured on our own GPUs.
9 of 9 models
Benchmarks, measured here
004.2
arc—light
Scores on the public suites, run on our own GPUs, so speed and quality are read on the same hardware.
Model
MMLU-Pro
GPQA
HumanEval+
MATH-500
First token
Arc-7
81.4
62.8
89.1
91.6
180 ms
Atlas 70B
74.2
51.3
80.4
82.0
160 ms
Forge Coder 34B
63.9
41.7
90.3
74.5
120 ms
Arc-7 Mini
68.5
44.0
78.2
79.9
70 ms
Forge Coder 7B
51.2
30.6
76.8
58.3
45 ms
Measured by Arclight in October 2026 on the same H200 nodes, five runs each, the median kept. First token at the 50th percentile, from Virginia.
Private models
Bring your own weights
Upload a model from Hugging Face, a bucket or a fine-tune you ran here. We compile it for the engine, serve it on the same endpoints and bill it by the second, with nobody else’s traffic on its GPUs.
Any architecture vLLM or TensorRT-LLM can run
LoRA adapters swapped per request, no reload
Private to your organisation, in the region you choose
From upload to endpoint in about eight minutes