Luminal
An open-source compiler that turns models into native GPU code, sold as serverless or on-prem.
- Founded
- 2025
- Example models
- GPT-OSS 120B, Llama 3 8B
By The Subconscious Team · Updated
What is 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.
The company sells that compiler two ways. Luminal Cloud is serverless inference in early access: upload a PyTorch or Hugging Face model, get a compiled endpoint with automatic batching and scale to zero, and pay per use. The on-prem license adds custom kernel work, dedicated engineers and SLAs. Luminal reports GPT-OSS 120B at 36K tokens per second on 8 H100s, against 28K for TensorRT-LLM and 26K for vLLM. Founders Joe Fioti, Jake Stevens and Matthew Gunton came from Intel, Apple and Amazon, went through Y Combinator in Summer 2025, and raised a $5.3M seed led by Felicis in November 2025.
Luminal pros, cons and use cases
Upsides
- Ahead-of-time compilation can beat runtime engines on throughput for the same model and hardware.
- Open-source compiler, so teams can inspect it or run it themselves before buying.
- Brings your own model, including custom architectures that hosted catalogs skip.
Core use cases
- High-throughput serving of a custom or fine-tuned model on owned GPUs.
- Teams that want an inference engine faster than vLLM without hand-writing kernels.
Downsides
- Very young company; the cloud is early access with no public price list or model catalog.
- Speed figures are self-reported aggregate throughput, not independent per-request benchmarks.
Luminal alternatives compared
Pick any row for the full head-to-head.
| Provider | Model access | Flagship models | Speed | Price | Customization | Deployment | Long context | Compare |
|---|---|---|---|---|---|---|---|---|
| Luminal | Bring your own weights | No public catalog | 36K tok/s on GPT-OSS 120B, 8xH100 (vendor) | Pay per use; rates not published | Compiles any PyTorch or HF model | Serverless (early access), on-prem license | Unknown | |
 Subconscious | Open weights | GLM 5.3, DeepSeek V4.1 Flash | 2x faster task completion | 50–80% lower cost; billed on processed tokens | Marathon post-trained variants | Managed API, dedicated, on-prem | 5M+ effective context | Compare |
 OpenAI | Closed, plus open gpt-oss | GPT-6 Astra, GPT-5.6 Sol, Terra, Luna | Fast mode: up to 2.5x at 2x price | $0.20–$10 in, $1.20–$50 out per 1M | N/A | API, Azure OpenAI, Bedrock | 1.05M; 2x input past 272K | Compare |
 Anthropic | Closed | Claude Fable 5.1, Opus, Sonnet, Haiku 4.5 | Fable is the slowest tier | $1–$10 in, $5–$50 out per 1M | N/A | API, Bedrock, Vertex AI, Microsoft Foundry | 1M, no surcharge past 200K | Compare |
 Google Vertex AI | Closed and open, 200+ models | Gemini 3.8 Flash, Claude, Gemma | Flash tier built for low latency | Gemini 3.8 Flash $0.75 in, $3.75 out | Custom training on GPUs or TPUs | Managed on Google Cloud | 1M on Gemini 3.8 Flash | Compare |
 Amazon Bedrock | Closed and open, 100+ models | Claude, GPT-6 Astra, Nova, DeepSeek | Latency-optimized option on some models | ~20–35% above direct; Claude at parity | Fine-tuning, Custom Model Import | Managed on AWS, AgentCore | Varies by model | Compare |
 Together AI | Open weights | Kimi K3, DeepSeek V4, GLM 5.2, Qwen 3.8 | 0.99s TTFT on DeepSeek V4 Pro | Parity with Fireworks and Baseten | LoRA and full SFT; RL in beta | Serverless, dedicated, GPU clusters | 512K on DeepSeek V4 Pro | Compare |
 Fireworks AI | Open weights | DeepSeek V4 Pro, Kimi K3 | 167–174 tok/s on DeepSeek V4 Pro | Fine-tunes served at base price | SFT, DPO, RFT; Training API | Serverless, dedicated GPUs | Full 1M on DeepSeek V4 Pro | Compare |
 Baseten | Open weights, 13 curated | GLM 5.2, DeepSeek V4, Kimi K3, gpt-oss 120B | 0.49s TTFT, lowest measured | H100 about $6.50/hr dedicated | Deploy any model with Truss | Model APIs, dedicated, self-host | Varies by model | Compare |
 Groq | Open weights | GPT-OSS 120B, Qwen 3.6 27B | 500–1,000 tok/s | Near the floor on small models | No fine-tuned model hosting | GroqCloud API | Around 131K max | Compare |
 Cerebras | Open weights | GPT-OSS 120B, Gemma 4 31B | ~3,000 tok/s on GPT-OSS 120B | $0.35 in, $0.75 out (GPT-OSS 120B) | Unknown | Shared API, dedicated, partners | Unknown | Compare |
 DeepInfra | Open weights | DeepSeek V4 Flash, Llama 3.1 8B | ~33 tok/s on DeepSeek V4 Pro (FP4) | From $0.02 per 1M | No managed fine-tuning | Shared API, no contracts | 66K on FP4 DeepSeek V4 Pro | Compare |
 Hugging Face Inference Providers | Open weights | GLM 5.3, Kimi K3, DeepSeek V4.1 Flash | Routes to fastest provider by default | Provider rates, no markup | N/A | Serverless router; dedicated Endpoints | Up to 1M, provider-dependent | Compare |
 Modal | Bring your own weights | None hosted | ~1s container boot | Per second; H100 $3.95/hr list | Run any training code | Serverless GPU containers | Depends on the model you deploy | Compare |
 Cloudflare Workers AI | Open weights | DeepSeek V4 Pro, GLM 5.3, Kimi K2.7 Code, gpt-oss 120B | Unknown | $0.011 per 1K Neurons; 10K free daily | BYO LoRA on small models (beta) | Serverless on Cloudflare network | 1M on DeepSeek V4; 262K on Kimi | Compare |
 xAI | Closed | Grok 4.6, Grok 4.20, grok-build | ~54 tok/s on Grok 4.6 | $2 in, $6 out (Grok 4.6); 2x past 200K | Unknown | First-party API | 500K (4.6), 1M (4.20, 4.3) | Compare |
 Mistral AI | Open weights, plus closed Codestral | Mistral Medium 3.5, Small 4, Large 3 | Unknown | $0.15–$1.50 in, $0.60–$7.50 out per 1M | Forge (enterprise); fine-tuning API deprecated | API, Azure, Bedrock, Vertex, self-host | 256K | Compare |
 DeepSeek | Open weights (MIT) | DeepSeek V4.1 Flash, V4 Pro | ~35 tok/s on V4 Pro | Off-peak hours at half price | Open weights to fine-tune | First-party API, Hugging Face weights | 1M, 384K max output | Compare |
 Moonshot AI | Open weights, custom license | Kimi K3, Kimi K2.6 | ~33 tok/s on Kimi K3 | $3 in, $15 out (Kimi K3) | Open weights to fine-tune | API, Kimi Code, OpenRouter | 1M | Compare |
 Z.ai | Open weights (MIT) | GLM-5.3, GLM-5.3-Flash | ~80 tok/s on GLM-5.3 | $1.40 in, $4.40 out (GLM-5.3); free Flash tier | Open weights, no license limits | API, GLM Coding Plan | 1M (GLM-5.3) | Compare |
 Alibaba Cloud | Closed Max; open smaller Qwen | Qwen 3.8-Max, Qwen 3.7-Max | ~40 tok/s on Qwen 3.8-Max | $2 in, $6 out international | No fine-tuning on Max | Model Studio on Alibaba Cloud | 1M (Qwen 3.8-Max) | Compare |
 Meta | Closed API; open Muse Glimmer | Muse Spark 1.3, Muse Glimmer | ~145–233 tok/s on Muse Spark 1.3 | $1.25 in, $4.25 out; Contributor tier cheaper | Open Muse Glimmer weights to fine-tune | Meta Model API (preview) | 1M | Compare |
 Cohere | Closed, plus open Command A+ | Command A+, Command A, Embed 4, Rerank 4 | 375 tok/s on Command A+ W4A4, per Cohere | $0.0375–$2.50 in, $0.15–$10 out per 1M | Enterprise fine-tuning, incl. private | API, Bedrock, Azure, OCI, VPC, on-prem | 256K on Command A; 128K on A+ | Compare |
 SambaNova | Open weights | MiniMax M2.7, GPT-OSS 120B, DeepSeek | ~820 tok/s on MiniMax M2.7 (SN50) | $0.22 in, $0.59 out (GPT-OSS 120B) | Unknown | SambaCloud, racks for neoclouds | Up to 192K (MiniMax M2.7) | Compare |
 Nebius | Open weights, 60+ models | DeepSeek, Qwen, GLM, Kimi, GPT-OSS | Among top hosts on throughput | From $0.06 per 1M input | Serve uploaded fine-tunes | Token Factory, dedicated, raw GPUs | Varies by model | Compare |
 Crusoe | Open weights | DeepSeek V4, GLM 5.3, Kimi K2.6, Nemotron 3 | Up to 9.9x faster TTFT vs vLLM (vendor claim) | $0.05–$1.74 in, $0.20–$4.40 out per 1M | Serverless LoRA fine-tuning | Serverless, self-serve and tailored dedicated, raw GPUs | Varies by model; cluster-wide KV cache | Compare |
 fal | Hosted media models | FLUX, Kling, Seedream | Cold starts on less popular endpoints | Per image, per video second, GPU time | LoRA training endpoints | Hosted API, serverless GPUs | Not applicable | Compare |
 Novita AI | Open weights | DeepSeek V4 Pro, Gemma 4 | ~36 tok/s on DeepSeek V4 Pro | From $0.02 per 1M; batch 50% off | Hot-swappable LoRA adapters | Serverless, GPU cloud, dedicated | Full 1M on DeepSeek V4 Pro | Compare |
 Venice | Open weights, plus proxied closed models | GLM 5.3, Kimi K3, DeepSeek V4 Pro | Unknown | $0.06–$12 in, $0.28–$60 out per 1M; DIEM staking | Unknown | Serverless API, consumer app | 1M on most current models | Compare |
 Parasail | Any Hugging Face model | GTE-Qwen2, Qwen3-VL-8B-Instruct | 600ms p99 real-time budget | Per-parameter rates; batch 50% off | Private Hugging Face repos | Serverless, elastic, dedicated, batch | Varies by model | Compare |
 Inference.net | Open, closed and custom | Customer fine-tunes | Batch windows of 24h to 7 days | Discounted spare GPU capacity | Distill traces into custom models | Batch API, gateway, dedicated GPUs | Varies by model | Compare |
 GMI Cloud | Open and third-party models | GLM-4.7-Flash, Google Veo | Near bare-metal performance | $0.07 in, $0.40 out (GLM-4.7-Flash) | Unknown | Shared, autoscaling, reserved GPUs | Varies by model | Compare |
| Thinking Machines | Open weights | Inkling, Inkling-Small | Unknown | Per 1M tokens by prefill, sample, train; Inkling $1.00 in, $4.05 out | LoRA SFT and RL via Tinker | Training API, beta serverless (Inkling only) | Inkling up to 1M; Tinker 32K–256K | Compare |
 Sail Research | Open weights | Kimi K2.6, GLM-5, GPT-OSS 120B | Minutes per turn by design | 30–80% off by completion window | Customer LoRA fine-tunes | API plus Sailboxes | Varies by model | Compare |
 Morph | Specialist models | morph-v3-fast, morph-v3-large | 10,500+ tok/s Fast Apply | ~40% fewer tokens than full rewrites | Fine-tuning offered | OpenAI-compatible API | Unknown | Compare |
 Relace | Specialist models | relace-apply-3, agentic search | ~10,000 tok/s apply | 3x+ cheaper than full rewrites | Unknown | Hosted API or self-hosted | 128K max | Compare |
 TypeSafe AI | Decision models | Jev, jev-1.13 | ~100ms per call | A fraction of an LLM call | Unknown | Early-access API | Unknown | Compare |
 StepFun | Open (Apache 2.0) and API models | Step 3.7 Flash, Step3 | ~128 tok/s on Step 3.7 Flash | $0.20 in, $1.15 out (Step 3.7 Flash) | Open weights to fine-tune | First-party API, OpenRouter | 256K | Compare |
 Runware | Hosted media models | Seedance 2.5, Qwen-Image-3.0 | Unknown | Images from fractions of a cent | Fine-tuned diffusion checkpoints | Unified API, raw GPUs | Not applicable | Compare |
 StreamLake | Proprietary coding models | KAT-Coder-Pro V2.5, KAT-Coder-Air | Unknown | Per token or KwaiKAT Coding Plan | Unknown | MaaS API, bare metal | Unknown | Compare |
 Wafer | Open weights | Qwen 3.5 397B Turbo, GLM 5.1 Turbo | 2–2.8x vs stock vLLM or SGLang | Wafer Pass from $10 a week | Agent-tuned dedicated deployments | Serverless pass, dedicated | Varies by model | Compare |
 RunInfra | Open weights | Nemotron 3.5 Lightning 30B, Qwen 3.8 27B | Cold starts under 2s | Coding plans from $10 a month | Uploads up to 50 GB; auto-quantization | Model APIs, agent-built endpoints | Varies by model | Compare |
 Particle.AI | Open weights | DeepSeek V4.1 Flash, GLM 5.3 Flash | ~157 tok/s on DeepSeek V4.1 Flash | $0.10 in, $0.40 out (GLM 5.3 Flash) | Unknown | Via Vercel AI Gateway | 1M | Compare |
 Infron | Closed and open, 400+ models | DeepSeek, Qwen, Claude, Gemini, GPT | Unknown | Provider rates; 3–5% top-up fee | Custom deployments | Gateway API, dedicated, BYOK | Varies by model | Compare |
Frequently asked questions
What is 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.
What is Luminal best for?
High-throughput serving of a custom or fine-tuned model on owned GPUs; Teams that want an inference engine faster than vLLM without hand-writing kernels.
How much does Luminal cost?
Luminal pricing at a glance: Pay per use; rates not published. Rates change often, so check Luminal's pricing page before committing.
What are the downsides of Luminal?
Very young company; the cloud is early access with no public price list or model catalog; Speed figures are self-reported aggregate throughput, not independent per-request benchmarks.
What are the best alternatives to Luminal?
Common alternatives include Subconscious, OpenAI, Anthropic, Google Vertex AI, Amazon Bedrock. Each has a head-to-head comparison with Luminal on this site.
Sources: Luminal, Luminal on GitHub, Luminal seed round, TechCrunch. Pricing and model lineups change often; figures are a snapshot.