Skip to content

// installation

Tested models

Applies to

Product: Server · Audience: Platform Operator

Models that basebox has actually verified, with metadata per entry: last tested, basebox version, backend, hardware, quantisation, context. This page lists only what has been verified in a documented way – not what should work (for that: Supported inference backends) and not what basebox recommends (for that: Recommended models).

Verified combinations

Model Quantisation Backend Hardware Context basebox Last verified Source
openai/gpt-oss-20b MXFP4 vLLM 0.15.0, external endpoint separate GPU host 32,768 1.7.1 2026 Deployment topologies
Language model of the reference configuration – vLLM 2 × H200 141 GB (1 GPU inference) up to about 258,000 input tokens in the test – Reference workload documented 2 × H200
Llama 3.3 70B Q8_0 · Q6_K · Q4_K_M – per VRAM of the customer server – – 2025-10-10 (Model Register) Model Register
OpenAI GPT-OSS 120B MXFP4 – per VRAM of the customer server – – 2026-05-05 (Model Register) Model Register
OpenAI GPT-OSS 20B MXFP4 – per VRAM of the customer server – – 2025-05-05 (Model Register) Model Register

The entries from the Model Register are the models shipped by basebox with their review dates from the compliance package; the register details are legal in nature (origin, licence, modification) and not a performance measurement.

Observed in testing (sizing reference)

The LLM recommendations contain observations from tests on the hardware classes, without basebox version and date being documented per row:

Observation Hardware Source
Qwen2.5 72B Int8 showed faster response times than Llama 3.3 70B FP8 4 × H100 LLM recommendations
Qwen3 4B Instruct FP16: ~23 tokens/s, practical maximum context ~30k 1 × 24 GB GPU LLM recommendations
Llama 3.3 70B AWQ may show high perplexity – Known issues
GPT-OSS 120B with data parallelism produces garbled output; use tensor parallelism – Known issues
Qwen 3 32B: only BNB-4bit with vLLM; append /no_think 4 × L40S LLM recommendations

These rows are sizing experience, not validation in the sense of the table above.

What an entry needs

For a model to appear here, basebox documents per combination:

  • Model (HuggingFace identifier), quantisation
  • Backend and version
  • Hardware (GPU model, count, TP), reference configuration
  • Context size and verified concurrency
  • basebox version, date
  • Verified features: chat, streaming, long context, tool calling (connectors), reasoning
  • Known limitations

You can report your own verifications to basebox in the same format – template under Using other models. That is how the list grows.

How to read the table

  • Verified means: in this combination basebox has confirmed function and behaviour. Other combinations of the same model (other GPU, other context) are compatible, not verified.
  • Context is the verified value, not the model's maximum.
  • Hardware is the test environment, not a requirement.
  • If a detail is missing, it is not documented – it is not estimated.

Next step: Recommended models