Open-source models, evaluated on your data.
Choosing a model is a business decision about quality, cost and risk. We shortlist by open-source licence, language and size, test candidates on your own documents, and size the hardware to the model you choose.
Model families we deploy.
We deploy only open-source models under Apache 2.0 or MIT licences. Sizes are total parameters, then parameters active per token for mixture-of-experts models. Figures reflect public releases as of October 2026 and change quickly, so we confirm versions in your proposal.
| Family | Size | Licence | Typical use |
|---|---|---|---|
| DeepSeek V4 | V4-Pro 1.6T, 49B active; V4-Flash 284B, 13B active | MIT | Coding, reasoning, price-performance, 1M-token context |
| GLM-5.2 | 753B total, 40B active | MIT | Agentic coding, long-context reasoning |
| Qwen 3.5 and 3.6 | 0.8B to 397B; 27B dense | Apache 2.0 | Multilingual work, single-GPU deployments |
| gpt-oss | 120b: 117B, 5.1B active; 20b: 21B, 3.6B active | Apache 2.0 | Reasoning on one 80 GB GPU (120b) or 16 GB (20b) |
| Mistral Small 4 and Large 3 | Small 4: 119B, about 6B active; Large 3: 675B, 41B active | Apache 2.0 | European languages, multimodal enterprise use |
We deploy only models released under permissive open-source licences, Apache 2.0 or MIT. Models under custom or restricted licences are not offered. We still confirm the licence of the exact release before deployment.
Which model fits which GPU.
Rule of thumb: weight memory is parameters times bytes per parameter, which is 2 for BF16, 1 for FP8 and about 0.5 for 4-bit formats. Add headroom for the context cache, which grows with context length and concurrent users.
| Model class | Examples | Approximate weight memory | Suitable systems |
|---|---|---|---|
| Compact, up to about 30B | gpt-oss-20b, Qwen 27B | About 16 GB (gpt-oss-20b) to 27 GB at FP8 | One RTX PRO 5000 or 6000, or L40S |
| Mid-size, about 100B to 120B | gpt-oss-120b, Mistral Small 4 | About 65 GB (gpt-oss-120b) to 120 GB at FP8 | One H100, H200 or RTX PRO 6000; two cards for headroom |
| Large, about 250B to 400B | DeepSeek V4-Flash, Qwen 3.5 (397B) | About 280 GB to 400 GB at FP8 | 8-GPU H200 or B200 server |
| Frontier, 600B and above | Mistral Large 3, GLM-5.2, DeepSeek V4-Pro | About 675 GB to 1.6 TB at FP8 | 8-GPU B300 server, or multiple nodes |
Estimates for planning only. Real requirements depend on quantisation, context length, batch size and the serving engine, which is why we benchmark on your workload before you buy.
How we choose and tune a model.
Four stages, each ending with results you can review.
Shortlist
Filter by open-source licence, languages, size and the hardware you can host.
Evaluate
Run candidates on your own test set, not public leaderboards, and compare quality, speed and cost.
Optimise
Quantise, fine-tune and add retrieval until the model meets your quality bar at your cost target.
Serve
Deploy with monitoring, versioning and a rollback path as new models arrive.
Serving engines
vLLM, SGLang and NVIDIA TensorRT-LLM, chosen per model and workload.
Quantisation
FP8 and 4-bit formats to cut memory and cost, with quality checked on your evaluation set before release.
Fine-tuning
LoRA and QLoRA adapters, full fine-tuning and distillation into smaller models, on data that stays on your infrastructure.
Retrieval and guardrails
Retrieval over your documents with access controls that follow your permissions, plus input and output filters.
Context length
Several current models support up to 1M tokens. Long contexts need much more memory, so we size for the context you actually use.
Licensing and control.
Every model we deploy is open-source licensed and runs on infrastructure you control.
Open-source licences only
Apache 2.0 and MIT allow commercial use, fine-tuning and redistribution. We confirm the exact release with you.
Runs on your infrastructure
Models run entirely on your own systems, on premises or in your chosen region.
Pinned and scanned
Weights are scanned and pinned to a version before use, with every change tracked.
Tell us which models you want to run.
Send the models, languages and number of users. We reply with the GPU configuration and an evaluation plan.