Tekindx embeds senior engineers inside your organisation to evaluate, fine-tune and run open-weight LLMs on infrastructure you control. Your first model is in production around week three, and you own the weights, the data and the roadmap.
Starts with a fixed-price working session. Your data never leaves your environment.
Artificial Analysis Intelligence Index, independent benchmark, retrieved 26 Oct 2026
| Model | Intelligence index | Speed | $ / 1M tokens |
|---|---|---|---|
| MiMo-V2.6-ProXiaomi | 46 | 57 t/s | $0.20 |
| DeepSeek V4.1 FlashDeepSeek | 45 | 169 t/s | $0.20 |
| GLM-5.3Z AI | 45 | 98 t/s | $0.90 |
| Step 5 PreviewStepFun | 44 | 229 t/s | $0.50 |
| GLM-5.3-FlashZ AI | 42 | 70 t/s | $0.10 |
| Muse Spark OpenMeta | 39 | 151 t/s | $0.80 |
| Qwen3.8 27BAlibaba | 34 | 45 t/s | $0.50 |
The shortlist follows your policy on model origin and licence. EU- and US-only builds are available.
Three senior engineers work inside your team from first workshop to handover. No model is chosen until it has been tested on your data.
One evaluation set.
Built from real work.
“Anything over 50k goes to a second reviewer.”
Approval depends on amount and region.
Resolved tickets, with the answer that was sent.
3 workstreams, signed off by your team
Illustrative example, not client data
Several open models.
One honest scorecard.
Same prompts, same cases, same hardware.
The API you would otherwise pay for.
“Under 800 ms, on the two GPUs we have.”
Scored on all 240 cases
Illustrative example, not client data
Weekly releases.
On your GPUs.
Adapters trained on outputs your team corrected.
Policies and precedents, cited in every answer.
Answer, escalate or refuse, by rule.
Adapter v3, citations required, rollback tested
Illustrative example, not client data
Your repo. Your runbooks.
Your next model.
Every change re-scored before it merges.
Monitor, roll back, upgrade.
Your engineers swap in next quarter’s model themselves.
Everything in your repositories
Illustrative example, not client data
Agents on open-weight models that read your systems, decide and write back. Dry run first, a person approves anything risky, and every step is traced.
Every step traced.
Nothing written without a rule or a person.
Supplier invoice arrives, no purchase order attached.
Search the ERP, read the contract, propose a posting.
Auto-post under 10k. A person approves anything above.
Invoice INV-20931, AED 14,200 · model: open weights, your GPUs
Illustrative example, not client data
Invoices, KYC packs, claims. Extracts fields against your schema, checks them against your records and files the result, with a confidence score and a link back to the source page.
Looks up the account, applies policy, takes the action (refund, re-issue, reschedule) and writes the note. Hands over to a person with the full context when it can't.
Reconciliations, PO matching, month-end exceptions. Runs on a schedule, proposes the journal, and posts only what your rules allow. Everything else lands in an approval queue.
On public benchmarks the best open-weight models trail the closed frontier by a margin that narrows every quarter, at a fraction of the price. On a single workflow, with fine-tuning on your own data, that gap usually closes. What changes is who controls the data, the cost and the upgrade path.
| Factor | Open weights, run by you | Closed API |
|---|---|---|
| Data | Stays in your cloud, data centre or sovereign region | Leaves your boundary for the provider's servers |
| Cost at scale | Fixed GPU cost that falls per request at sustained volume; we show you the break-even before you commit | Per-token pricing that rises with usage |
| Customisation | Fine-tune on your data and keep the weights | Prompting, plus limited tuning on the provider's terms |
| Model choice | Swap between Mistral, Qwen, DeepSeek, Inkling and others as they improve | One vendor's roadmap |
| Continuity | The version you validated keeps running until you change it | Models deprecated on the provider's schedule, often with months of notice |
We deploy wherever your risk team needs the model to live, and we document every control so procurement and audit have what they need.
Inside your own VPC on AWS, Azure, Google Cloud or a sovereign cloud region. No traffic to model vendors.
Fits most regulated enterprises
On your own NVIDIA or AMD GPU clusters, integrated with existing identity, logging and change control.
Fits banks, telcos and healthcare
Fully disconnected environments, with weights transferred and verified through your approved media process.
Fits defence and critical infrastructure
Every engagement starts with a fixed-price working session, so you know the scope and the price before you commit.
Enterprises and government
For organisations that can't send sensitive data to a third-party API but need AI in production this year.
GPU clouds, model labs and integrators
For partners whose customers stall between pilot and production and need senior engineers to get them there.
Everything lives in your repositories from day one. When a better model ships, your team can test it and swap it in.
The short answers. The engineer you meet will go deeper on any of them.
No. The work happens inside your environment, under your own access controls. Data, prompts and model weights stay in your cloud account, sovereign region or data centre.
Yes. We run on GPUs in your cloud account, in a sovereign region or on-premises, and we size the hardware during the working session, with the break-even against per-token APIs before you commit.
Every engagement starts with a fixed-price working session, so scope and price are agreed before anything else is committed. Builds then run eight to sixteen weeks.
Your first model is usually in production around week three, on one workflow, with an evaluation set your team has signed off.
We test the shortlist against your current approach or a closed model on your own cases, and show you the scorecard before you build. You decide with the numbers in front of you.
Mistral, Qwen, DeepSeek, Kimi, GLM, Llama, Gemma and others, chosen by evaluation on your data and filtered by your policy on model origin and licence.
Everything: the weights and adapters, the serving code, the evaluation set and its CI job, and the runbooks to monitor, roll back and upgrade. Your team runs it without us.
Start with a five-minute conversation with our discovery agent, by voice or text. The engineers you meet next read the transcript, so you never explain the problem twice.
A five-minute conversation with our discovery agent. Speak or type, in your own language. Stop whenever you like.
You're talking to an AI agent. The conversation is transcribed and, if you speak, recorded, so a Tekindx engineer can review it.
You're verified. How would you like to talk?
Our discovery line opens shortly. Until then, email an engineer directly with your company, the workflow you have in mind and your timeline. We reply within one business day.
Thanks for your time. If there's a fit, you'll have a calendar invite in your inbox. Otherwise we'll email you within one business day.
Same content as the human view, structured for parsing. Link straight to this view with #machine.