Tested product boundary
Compatibility Matrix
A precise separation between verified public evidence, supporting transfer evidence, development coverage, and unverified roadmap items.
How to read this page
Compatibility is evidence-scoped.
A verified row means that the listed configuration appears in the disclosed completed study. It does not mean every configuration inside that category will produce the same speed or quality result.
Optimizer and workload
Current evidence boundary.
| Dimension | Configuration | Status | Evidence boundary |
|---|---|---|---|
| Optimizer | AdamW | Verified | Matrix54, RC1.06 three-pass endurance, and full-parameter transfer. |
| Optimizer | Muon | Verified | Matrix54, RC1.06 three-pass endurance, and Matrix12-v3. |
| Optimizer | Lion | Verified | Matrix54, RC1.06 three-pass endurance, and Matrix12-v3; degraded rescues carry an explicit caveat. |
| Optimizer | Other optimizer families | Qualification | Adapter and fallback behavior must be validated before a performance test. |
| Training phase | Continued pretraining | Primary | Primary public evidence across all disclosed studies. |
| Training phase | Pretraining and SFT | Development | Internal research coverage only; no headline public performance claim. |
| Adaptation | LoRA | Primary | Matrix54 at 3B and three-pass endurance across AdamW, Muon, and Lion at 1.5B. |
| Adaptation | Full parameter | Supporting | Matrix12-v3, 1.5B, one development seed. |
| Adaptation | QLoRA | Development | Research coverage exists, but it is outside the frozen public headline studies. |
Scale and runtime
Hardware and systems boundary.
| Dimension | Configuration | Status | Notes |
|---|---|---|---|
| Model family | Qwen2.5 | Verified | 1.5B and 3B sizes in the disclosed evidence. |
| Model scale | 1.5B and 3B | Verified | Results remain protocol-specific. |
| Model scale | Above 3B | Qualification | No public headline claim; memory and runtime paths require a customer-specific gate. |
| Hardware | NVIDIA CUDA GPU | Verified | Single-GPU internal evidence. |
| Framework | Linux and PyTorch | Verified | Exact versions are frozen with each release artifact. |
| Distributed | FSDP / DeepSpeed | Planned | Architecturally adjacent, but not verified in the disclosed public matrix. |
| Hardware | AMD ROCm | Planned | Portable fallback is a design goal; no current verified performance claim. |
| External accelerator | Unsloth | Research only | Exploratory work does not constitute a current pilot integration or performance claim. |
Customer acceptance rule
A configuration enters a paid performance gate only after the exact package, fallback path, checkpoint behavior, and matched baseline pass compatibility qualification.