Get a free optimization plan for your model
Tell us what you run and where. Within two business days you get a written plan that says which quantization format to use, how much memory and money it should save, and how to prove quality held up.
Your plan includes
- A format recommendation — for example FP8, AWQ 4-bit or GGUF Q5_K_M — matched to your hardware and serving runtime.
- Memory and throughput estimates for weights, KV cache and batch size, worked out for your context length.
- A cost comparison of your current setup against the recommended one, with the assumptions written down.
- A quality-check plan: which metrics and test prompts to run so you can ship with evidence, not hope.
- The exact tooling steps to produce and serve the quantized model.
Quick estimate before you start
Could save about — a month (— a year).
An illustration only. The plan replaces this with numbers for your workload.
No sales call required. Never paste secrets, API keys or private data — we work from public model information and the details above.
What happens next
You send the form
Two minutes, no account. The more context you give, the sharper the plan.
We work out the numbers
Formats, memory, throughput ceilings and costs for your hardware and runtime, using the same methods as our public calculators.
You get the plan by email
Usually within two business days, as a short written document you can share with your team.
Optional hands-on help
If you want someone to implement it, we can introduce an engineer from our network with a fixed quote up front. Entirely optional.
An excerpt from an example plan
Illustrative example for a hypothetical team, not a client result.
Setup: Llama 3.1 70B, BF16, vLLM on 4 × 80 GB GPUs, 8K context, about 30 concurrent requests.
Recommendation: FP8 weights (W8A8) with FP8 KV cache. Weights drop from about 131 GB to about 66 GB, which fits on 2 GPUs with room for the KV cache of 30 × 8K-token sequences (about 37 GB at 8-bit).
Quality check: run your 200 most common production prompts through both versions and compare with a task-specific score and side-by-side review before switching traffic.
Expected effect: half the GPUs for the same concurrency, subject to the latency test in step 4 of the plan.
Numbers computed with the method in our VRAM calculator and KV cache guide.
Prefer to start with a cheat sheet?
The one-page reference to every formula and format on this site, ready to print. We will email it to you and add you to the weekly explainer.
Questions
Is the plan really free?
Yes. It is how we learn what teams need and how some of them come to hire help later. There is no obligation and no automatic follow-up sequence.
Do you need access to our model or data?
No. We work from public model configurations and what you tell us. Please do not send credentials, proprietary weights or customer data.
Are your recommendations tied to a vendor?
No. We recommend whatever format and runtime fits your situation. If a recommendation ever involves a partner we earn a referral fee from, the plan says so clearly.
What if the model is a fine-tune or a custom architecture?
Tell us the base model and what changed. Most fine-tunes quantize like their base; custom architectures may need a calibration step, which the plan will describe.
Is our request confidential?
Yes. Details you send are used only to prepare your plan and reply to you, as described in our privacy policy.