vs
Luminal vs Boundless
Luminal compiles models into native GPU kernels and licenses its compiler on-prem. Boundless hosts a 22-model catalog and tunes reserved capacity for you.
By The Subconscious Team · Updated
Luminal vs Boundless: key differences
Luminal's open-source compiler turns PyTorch or Hugging Face models into native kernels ahead of time, and Luminal reports GPT-OSS 120B at 36K tokens per second of aggregate throughput on 8 H100s against 26K for vLLM. Luminal Cloud is in early access, with DeepSeek V4.1 Flash at $0.1425 in and $0.57 out per million tokens, and an on-prem license adds custom kernel work, dedicated engineers and SLAs. Boundless charges $0.20 in and $1.00 out for DeepSeek V4.1 Flash on an API that works with the OpenAI and Anthropic SDKs and also carries GPT and Claude. Its tailored inference has Boundless engineers tune reserved capacity against uptime, throughput and latency targets in the contract, and it publishes no benchmarks of its own.
Pick Luminal to serve your own or custom-architecture model, on your own GPUs if needed, or for the lower DeepSeek V4.1 Flash price. Pick Boundless for a hosted catalog with long context and closed models on one key, or for managed capacity you can test free for two weeks on real traffic.
What Luminal and Boundless do
Luminal
Luminal builds an inference compiler. Where vLLM and SGLang interpret a model at runtime, Luminal compiles it ahead of time into native kernels for GPUs and ASICs. Models get lowered to a small graph of 15 primitive ops, and the compiler searches over fusion, tiling, memory and scheduling choices instead of relying on hand-written rules, which it says can find optimizations like Flash Attention on its own. The compiler is open source in Rust under Apache 2.0 or MIT, runs on CUDA and Metal with ROCm on the roadmap, and works as a torch.compile backend.
Example models: DeepSeek V4.1 Flash, Ideogram 4.5
Full Luminal profileBoundless
Boundless launched in September 2026 as an inference partner for AI-native startups. It sells two things. The Model API serves pre-tuned models on shared capacity, billed per token from prepaid credits with a budget per API key. Tailored inference reserves capacity for one customer, and Boundless engineers pick the model on the customer’s own evals, tune precision, batching, caching, speculative decoding and parallelism, then run the deployment against uptime, throughput and latency targets written into the contract. A free two-week evaluation tests those targets on real traffic against the customer’s current provider first.
Example models: GLM-5.3, DeepSeek V4.1 Flash, Kimi K3
Full Boundless profileShould you choose Luminal or Boundless?
Luminal vs Boundless at a glance
| Attribute | Luminal | Boundless |
|---|---|---|
| Model access | Small hosted catalog plus bring your own weights | Open weights, plus closed GPT and Claude |
| Flagship models | DeepSeek V4.1 Flash, Ideogram 4.5 | GLM-5.3, DeepSeek V4.1 Flash, Kimi K3 |
| Speed | 36K tok/s on GPT-OSS 120B, 8xH100 (vendor) | 342 tok/s on DeepSeek V4.1 Flash 1017 (Vercel) |
| Price | $0.1425 in, $0.57 out (DeepSeek V4.1 Flash) | $0.04–$2.30 in, $0.21–$11.40 out per 1M (open) |
| Customization | Compiles any PyTorch or HF model | Serving-stack tuning, no fine-tuning |
| Deployment | Serverless (early access), on-prem license | Serverless API, reserved tailored capacity |
| Long context | Unknown | 1M context on most models |
Frequently asked questions
What is the difference between Luminal and Boundless?
Luminal compiles models into native GPU kernels and licenses its compiler on-prem. Boundless hosts a 22-model catalog and tunes reserved capacity for you.
When should I choose Luminal over Boundless?
Compiling your own models; On-prem deployment with SLAs; Cheaper DeepSeek V4.1 Flash.
When should I choose Boundless over Luminal?
A 22-model hosted catalog; GPT and Claude beside open models; Free two-week evaluation.
Is Luminal or Boundless cheaper?
Luminal: $0.1425 in, $0.57 out (DeepSeek V4.1 Flash). Boundless: $0.04–$2.30 in, $0.21–$11.40 out per 1M (open). The cheaper choice depends on the model and workload.
Related comparisons
Running long-horizon agents?
If your agents run past 200K tokens, compare both against Subconscious. Our inference stack treats a long-horizon trace as the primary workload, so speed, cost, and accuracy hold up deep into the trace.