A KVP white paper — drawn from a live chemical business RFQ deployment on an open-weighted model — on when to run your own LLM, when to buy a frontier API, and how to govern both safely.
The industry framing of "open vs closed" is misleading. In practice they solve different enterprise problems and increasingly ship side-by-side in the same stack.
Weights are published under permissive licences. You download, fine-tune, quantise, deploy — on your GPUs, in your cloud, on-prem, air-gapped if needed.
Closed weights, API-only. Trained on far larger compute budgets with proprietary post-training. Lead the public benchmarks on reasoning, agentic tool use and long-horizon tasks.
| Dimension | Open-weighted | Frontier |
|---|---|---|
| Data sovereignty | Full — weights + data stay in your VPC | API traffic leaves the tenant boundary |
| Peak reasoning | Good, closing the gap fast | Best available today |
| Fine-tuning on proprietary data | Native — LoRA, QLoRA, full-parameter | Limited to hosted fine-tune APIs |
| Inference cost at scale | $0.20–$0.60 per 1M tokens (self-hosted) | $3–$50 per 1M tokens (API) |
| Latency control | Deterministic — tuned to your GPUs | Vendor-dependent, variable |
| Time to first pilot | 4–8 weeks (infra + fine-tune) | Days (API key + prompt) |
| Long-horizon agentic work | Emerging — DeepSeek R2, Qwen 3 | Mature — Fable, GPT-5, Gemini 2.5 |
| Vendor lock-in | None — swap models freely | High — prompt & tool schema coupling |
| Compliance / audit | You own logs, evals, red-team runs | Depends on vendor's DPA and ZDR terms |
| Talent needed | MLOps + GPU ops + fine-tuning skill | Prompt engineering + integration |
A large, diversified chemical business — annual revenue in the multi-billion range, operations spanning 30+ countries, and a procurement catalogue of 10,000-plus materials, equipment, spares and services — was drowning in inbound RFQs. The tendering process is highly controlled and process-governed: each procurement category follows its own multi-stage workflow, from technical and commercial evaluation to committee approval and audit trails. The team actively manages procurement worth over ₹2,000 crore per year and processes 400+ live tenders a month across categories as varied as raw materials and additives, capital machinery, spares, civil and electrical contracts, labour contractors, safety and logistics services, and office stationery. Every day buyers received hundreds of PDFs, emails and portal exports, each mixing free-text specifications, drawings, quantities, delivery terms and compliance requirements. Turnaround was averaging three business days; the slow response meant production schedules were at risk and the best-priced or technically qualified suppliers were often missed because faster competitors locked in the order first.
Started with Qwen 3 32B for the balance of reasoning quality and inference footprint. LoRA fine-tune on 4 years of historical tender → awarded-contract pairs, followed by DPO alignment on buyer-approved responses.
Deployed on 4× H200 nodes inside the customer's data centre. vLLM + TensorRT-LLM for inference, quantised to FP8. Peak throughput 1,400 concurrent tokens/sec at P99 latency < 900ms.
10,000-plus materials, spares, equipment and services plus datasheets embedded with BGE-M3, stored in a private vector database. The model reasons over the RFQ, retrieves the top candidate items, and grounds its recommendation in citations.
Every model draft goes to a buyer with confidence scores, retrieved evidence and a redline diff. Rejections feed the next fine-tune cycle. No autonomous award or contract to external vendors.
RFQ turnaround time
Buyer productivity uplift
Award-rate on negotiated RFQs
Data retained inside customer VPC
Anonymised per KVP's client confidentiality policy. Large multi-billion-revenue diversified chemical business, procurement catalogue > 10k line items, deployment 2026.
Score the workload on each lens. If three or more push toward one side, that is your model family. If it's a split, the answer is a portfolio — route each request to the right model.
Open-weighted: Regulated, sovereign, export-controlled or IP-critical data → open-weighted, private deployment.
Frontier: Public, marketing, general-purpose content → frontier via API with ZDR is fine.
Open-weighted: Narrow, repeatable tasks — extraction, classification, routing, drafting → open-weighted, fine-tuned, wins.
Frontier: Long-horizon reasoning, agentic tool use, rare cross-domain skill → frontier still leads.
Open-weighted: Sustained > 500M tokens/month or predictable batch inference → self-hosted open beats API.
Frontier: Spiky, low-volume, exploratory workloads → frontier API is cheaper end-to-end.
Open-weighted: You have 6–12 weeks and MLOps talent → build a fine-tuned open-weighted asset that compounds.
Frontier: You need a pilot in the next sprint → start on frontier, measure, then decide.
Take 200 real production tasks. Score frontier vs a fine-tuned open model on your own evals — quality, latency, cost, safety. Decisions on benchmarks alone are almost always wrong.
Classify the workload's data before you compare models. Once the classification is 'confidential' or higher, the shortlist collapses to open-weighted, private-deployed options.
Wrap every LLM call behind an internal gateway with a stable schema. You will swap models — plan for it. This is what makes a portfolio strategy actually workable.
Both families hallucinate. Ship with reviewers on all outbound content until eval scores plateau across three release cycles — then thin the loop, don't remove it.
Frontier API pilots look cheap. Multiply pilot volume by 20× for production, add growth, and re-check. Open-weighted capex is front-loaded but flattens the curve fast.
Prompt, retrieval, tool call, response, evaluator score, human decision. Open or frontier, the audit trail is what turns AI risk from unknowable into managed.
One page: what data classes may use which models, who approves fine-tunes, how outputs are reviewed. This ends the shadow-AI problem faster than any tool.
Every serious enterprise AI estate KVP is building in 2026 uses both families. Frontier models sit at the top of the stack for reasoning-heavy, agentic, cross-domain work — Agentforce, autonomous coding, complex customer-service triage. Open-weighted models sit underneath for the volume workload — extraction, classification, drafting, retrieval-grounded answers on proprietary data.
The winning pattern is a routing layer that sends each request to the right model based on data class, task complexity and cost budget. That routing intelligence — the evals, the prompts, the guardrails, the fine-tuned open models — is the durable IP. The model itself is replaceable.
Enterprises that pick a religion (all-frontier or all-open) will lose. Enterprises that build a portfolio and the routing muscle to run it will compound faster than either camp alone.
A fixed-scope engagement to design and stand up your open-weighted + frontier portfolio — model bake-off, private deployment, routing layer, evals and governance policy.
Score frontier vs fine-tuned open on 200 real tasks. Data-class the workload.
Stand up open-weighted inference in your VPC / on-prem, fine-tune on proprietary data, wire retrieval.
Model gateway, per-request routing rules, eval harness, internal model policy handover.
What a frontier autonomous engineer means for Salesforce development, SI value and enterprise ROI.
ReadConnecting Salesforce to Anthropic via MCP — five use cases that pay back in 90 days.
ReadGet Salesforce AI-ready before you layer on Agentforce — KVP's six-step cleanup playbook.
Read