Low
Qwen3.5-9B
High-volume sorting, routing and pulling data out of forms. Fits on a single low-cost GPU.
The easy majorityLarge models to small models
Cut your AI bill.
Move routine work
to smaller models.
ModelShift measures where your LLM spend goes, tests open-weight models, caching and routing on your real workloads, and only ships what holds up. It starts with a free 15-minute AI cost review.
- 85–90%
- of AI traffic is routine*
- 97.5%
- quality bar before shipping
- 15 min
- free review, no system access
requests
small models
frontier
*Our working assumption. The audit measures your real split.
Models and open tooling we test with
Qwen3.5-9BQwen3.8-27BGLM-5.3-FlashDeepSeek-V4.1-FlashvLLMUnslothPromptfooGarakMicrosoft PresidioLlama Guard 3AWS BedrockGoogle Vertex AIRunpodBasetenFireworksOpenRouter
The problem
Most AI traffic is routine. You pay frontier prices for all of it.
Classifying tickets, pulling fields out of documents, answering from internal docs, writing SQL. A smaller model trained on your own examples can often do this work faster and for less, while the genuinely hard requests still go to the big model.
87 routinetickets, extraction, internal Q&A, SQL
13 hardnovel reasoning, long agent chains
Cost
Every request pays frontier per-token pricing, including the easy ones. We count the full cost, hosting included, so smaller never gets assumed to mean cheaper.
Speed
Large models are slower to first token. On structured tasks a tuned small model can answer before the big one starts.
Control
Your prompts and customer data leave your perimeter on every call. A small model can run inside your own cloud account.
How it works
Explore, test, build, ship. In that order.
Each step earns the next. If the numbers don't justify building anything, we tell you and you've lost fifteen minutes.
-
01 / Explore
Free 15‑minute AI cost review
Tell us which AI tools you use, roughly what they cost and what you most want fixed: cost, speed, security or reliability. You get first recommendations on the call.
Free -
02 / Test
Paid technical audit
We baseline cost, quality and latency on representative, permission-approved data, then test caching, routing and smaller models against it. You get the evidence, the risks and the full running cost.
Fixed scope, quoted after the review -
03 / Build
Proof of concept or production build
Fine-tune a small model on your workload, add a router in front of your frontier API, or automate a manual workflow with a human approval step. Only where the audit says it pays.
Only if the audit supports it -
04 / Ship
Deploy, monitor, improve
Run it in your own cloud or on a dedicated managed endpoint, with drift alerts, a monthly cost-and-savings report and model upgrades as better open weights are released.
Optional ongoing care
The router
A small model handles the routine. Three checks catch the rest.
A gateway sits between your app and your models. Easy, familiar requests are answered locally. Anything long, unusual, low-confidence or malformed is sent on to your existing frontier model before your user ever sees it.
Your appOpenAI-compatible
Frontier modelyour existing contract
Small modelin your cloud
Gate 1pre-flight
Gate 2confidence
Gate 3schema
answered locally
escalated
of all-frontier token cost*
*Illustrative, assuming a local request costs about a tenth of a frontier one before hosting. The audit replaces every assumption here with your measured numbers.
Before the small model runs · ~12 ms
Gate 1: pre-flight
Very long prompts, or ones that look unlike anything in your training data, skip straight to the frontier model so no GPU time is wasted.
First 30 tokens, held back
Gate 2: confidence
The opening tokens are checked for calibrated confidence before anything streams to the user. Low confidence means a silent hand-off.
Before the response returns
Gate 3: schema
JSON, SQL and tool calls are validated against your schema. Anything invalid is retried on the frontier model automatically.
The pipeline
Six stages from your logs to a model you own.
Built from standard, battle-tested open-source parts. No custom inference forks, no black boxes, and a human sign-off before anything reaches production.
- 1
Qualify
We understand the workflow, your AI spend, latency needs and data rules, and work out the full cost of running anything new, GPUs included.
- 2
Benchmark
Permission-approved samples are scrubbed of personal data inside your environment. 20% is locked away as a blind holdout. Candidate models are scored by programmatic checks and an independent judge from a different model family.
Presidio + spaCy scrubber20% blind holdoutCross-family judge - 3
Adapt
Where a small model gets close, we fine-tune it on your hard cases and a replay of everyday traffic so it doesn't forget the easy ones, then calibrate its confidence.
Unsloth LoRADPOFP8 / AWQ - 4
Secure
A 500-probe red-team mapped to the OWASP Top 10 for LLMs, a safety sidecar in the gateway, and a re-check that legitimate requests aren't being refused.
GarakPromptfooLlama Guard 3 - 5
Deploy
A single Helm chart into your Kubernetes cluster, or a dedicated managed endpoint. You own the fine-tuned weights either way.
vLLMEKS / GKE / AKSBaseten / Fireworks - 6
Monitor
Drift alerts when escalations climb, a monthly report of tokens served locally against spend avoided, and retraining every 60 to 90 days or when drift is detected.
Models
Right-size the model to the task.
Our starting shortlist, all open-weight. Nothing is recommended until it has been benchmarked on your workload with full running costs.
Medium
Qwen3.8-27B
Reading PDFs, answering from internal documents, everyday coding, vision and OCR.
General business assistantsHigh
GLM-5.3-Flash
Multi-step agents and very large document batches, via API from day one.
Heavier reasoningEscalation
Your frontier model
The hardest 10 to 15% of requests, through the contract you already have.
Kept for what needs itTrack A
In your own cloud
For teams with strict InfoSec rules and existing cloud credits.
- Deploys inside your AWS, GCP or Azure account with one Helm chart
- Inference never leaves your perimeter
- You pay your cloud provider for GPUs directly
- High availability across two zones
Track B
Managed endpoint
For fast-moving teams that want the saving without an MLOps hire.
- Dedicated single-tenant GPUs, US or EU, zero data retention
- No infrastructure for you to run
- Switch over by changing two lines of config:
# The only change in your app client = OpenAI( base_url=os.environ["MS_URL"], api_key=os.environ["MS_KEY"], )
Savings
Estimate what shifting could save you.
A rough first look based on our routing design. The free review sharpens it; the audit measures it.
Everything you pay OpenAI, Anthropic, Google and similar each month.
Classification, extraction, internal Q&A, SQL and similar bounded tasks.
Estimated saving per month
£0
£0 a year, 0% of today's spend
Frontier, escalatedSmall-model hosting
Want this checked against your real usage? Leave your email and we'll come back with a personal breakdown.
How we estimate: the routine share, less a small allowance for requests the gates escalate, moves to a small model; everything else stays on your frontier API at today's price. Hosting assumes two GPUs across two zones for high availability, around £1,900 a month in your cloud or £2,000 managed, rising by half for real-time latency. Excludes ModelShift fees, which are quoted after the audit. Indicative only, not a quote or a promise of savings.
Who it's for
Built for teams with a real, recurring AI bill.
We work best with companies already paying for LLM APIs on production workloads, and with the people who own that number.
AI SaaS teams
Your margin depends on inference cost per user.
AI agencies
You resell AI workflows and want a better unit cost on every client.
Engineering and platform leads
You own the LLM bill and the latency targets.
Finance and FinOps
You need a number you can defend, not a vendor promise.
Evidence before invoices
No savings, performance or compliance claims without a reproducible test on your workload.
Your data, your rules
Only authorised data, scrubbed in your environment first, with zero data retention wherever a third party is involved.
You own what we build
The fine-tuned weights, the validators and the deployment are yours. No lock-in to us.
The book
We're writing the book on it.
By Joshua Boladuro and David Willis Owen
The Great Model Shift is a research-led guide to when smaller or open-weight models can replace larger ones, backed by published studies, interviews with researchers and benchmarks you can reproduce.
- The rise of large language models and the business problem
- Open-weight and small models: what they are and when they fit
- Auditing AI usage and calculating the full cost
- Choosing models and testing quality
- Distillation, fine-tuning and quantisation
- Deployment, security and running costs
- Real experiments: savings, trade-offs and failures
- Turning an audit into a useful client service
The team
Two engineers, one question: does it need the big model?
ModelShift pairs hands-on AI engineering with security and deployment depth. We run the free reviews ourselves.
Joshua Boladuro
Co-founder · AI engineering and growth
AI engineer who builds production AI assistants and workflow automation. Leads the cost reviews, client conversations and the shift from audit to build.
David Willis Owen
Co-founder · security and model deployment
Leads benchmarking, fine-tuning and the inference pipeline, with a security-first view of data handling. Publishes our model cost and speed benchmarks.
Benchmarks you can rerun, talks you can question.
We publish model cost and speed teardowns on YouTube, LinkedIn and X, and speak at AI events in London. Want us at yours, or on your podcast? Use the contact form below.
YouTubeLinkedInXLondon AI events
Free review
Book your free 15‑minute AI cost review.
Costs climbing, slow responses, data leaving your cloud, a manual process you want automated. Bring whatever's hurting most. You leave with one or two concrete improvements, whether or not you ever work with us.
Review requested.
Thanks. We'll email you within one working day with times for your 15-minute call. In the meantime, try the savings estimate with your real numbers.
FAQ
Questions we get on every call.
Do you need access to our systems for the free review?
No. The free review is a conversation about your tools, rough spend and the problem you want fixed. We never ask for passwords, API keys or customer records.
Will a smaller model really be as good?
Only on the tasks where testing says so. We grade candidates against a blind holdout of your own data that's never used for training, and our design target before anything ships is 97.5% or better against your current outputs.
Anything the small model isn't sure about is escalated to your frontier model automatically.
What if our workload isn't a good fit?
We tell you. Mostly open-ended creative work, or spend below a few thousand pounds a month, usually won't justify hosting a model. Caching, prompt trimming or smarter routing often will, and we'll say which.
Is the review only about cost?
No. Cost is where most teams start, but latency, data security, reliability and manual workflows all come up. Tell us what's hurting and we'll focus the call on that.
Who owns the fine-tuned model?
You do. The adapted weights, validators and deployment are handed over to you, whether the model runs in your cloud or on a managed endpoint.
Where is our data processed?
Personal data is scrubbed inside your environment before anything is shared with us. In your-cloud deployments, inference stays in your account. Managed endpoints and any third-party models we use run in US or EU regions with zero data retention.
How much does the paid audit cost?
It depends on the number of workflows and the data involved, so we quote a fixed scope after the free review. We'd rather start with a small, clearly scoped job than a large migration on day one.
Find out what your AI bill should be.
Fifteen minutes, three questions, no system access. If there's nothing worth shifting, we'll tell you that too.