Model Porting Analysis

This analysis reflects the public model landscape in August 2026. A model is only an exact port when its tensor layout and computation are reproduced; sharing a flow-matching objective is not enough.

Target	Fit with FlowEdge today	Missing exact-port work	Recommendation
Custom encoder + FlowEdge MLP flow head	Direct	Train/export flow.*; feed one F32 condition vector	Production-ready integration path
Mamba-1 language/control checkpoints	Strong backbone fit	LM projection, sampling, tokenizer adapter, checkpoint mapping	Small, useful port
Mamba-3 SISO/MIMO	Strong architectural fit, incompatible recurrence	Complex/angle state, trapezoidal update, K/V state, grouped heads, LM head	Deferred behind SmolVLA/Tenstorrent
SmolVLA	Partial external-cache path	Transformer flow expert with self/cross attention over VLM tokens	Validate captured VLM-cache replay, then port encoder
pi0 / pi0.5	Objective fits; architecture does not	PaliGemma/VLM token interface and coupled flow action expert	Adapter first, exact port later
GR00T N1.5	Objective fits; architecture does not	DiT action model cross-attending to Eagle VLM embeddings	External runtime first
pi0-FAST	Does not use the flow path	FAST action tokenizer and autoregressive transformer decoder	Defer unless token actions become a core goal

What works immediately

The head-only checkpoint and sample_condition ABI are intentionally model-runtime neutral. A company can keep a supported VLM or perception encoder in PyTorch, JAX, TensorRT, or ONNX Runtime, train the compact FlowEdge velocity MLP against its pooled latent, and deploy only that action head in this engine. This is an integration or distillation target, not a claim that an upstream checkpoint converts unchanged.

This boundary is useful because current VLA action experts are structurally richer than FlowEdge’s MLP. SmolVLA uses a transformer flow-matching expert with interleaved attention. OpenPI describes pi0 and pi0.5 as flow-based models, while pi0-FAST is autoregressive. GR00T N1.5 uses a DiT that cross-attends to VLM embeddings. Their weights cannot be flattened into flow.cond_proj without retraining.

Deferred backbone option: Mamba-3

The active priority is SmolVLA’s exact action expert and Tenstorrent execution, as specified in the roadmap. Mamba-3 is a later option, starting with SISO decode and then MIMO. The official Mamba repository now includes Mamba-3, and the paper describes it as inference-first, with a richer recurrence, complex-valued state updates, and a MIMO formulation. Those properties align with FlowEdge’s fixed state, streaming-first runtime better than adding a general transformer graph executor.

The implementation should be a new models/mamba3/ module rather than conditionals in Mamba-1:

1. Add scalar reference kernels for SISO recurrence and step decode.

2. Add AVX2/NEON state-major kernels, keeping angle and recurrent state caller-owned.

3. Define the Mamba-3 state payload and architecture identifier inside the existing versioned, checksummed snapshot envelope.

4. Add embedding, normalization, and LM output projection, with top-k/top-p sampling outside the deterministic backbone call.

5. Validate token-by-token output and state against the official implementation before optimizing chunked prefill.

6. Add MIMO only after SISO correctness; its extra K/V state should have an explicit arena formula.

Mamba-2 remains useful as an intermediate reference, but implementing it first only makes sense if a specific customer checkpoint requires it. Neither additional Mamba port takes priority over the active CUDA then SmolVLA robotics program.

Recommended VLA port boundary

Exact SmolVLA, pi0, and GR00T ports need a condition-sequence ABI, not just a larger vector. The Transformer embedding API starts that boundary with a read-only matrix [tokens, width] and an optional prefix attention mask. Completing the contract still requires a lifetime valid across resumable solver calls and action-expert cross-attention. The ABI can serve several action experts without importing every VLM into Core.

SmolVLA is the best first exact VLA expert because it is compact, openly documented, and explicitly designed for low-latency inference. pi0/pi0.5 should follow through an OpenPI conformance harness. GR00T should initially remain an external NVIDIA runtime feeding an independently trained FlowEdge head; reproducing Eagle plus DiT would turn this project into the broad graph runtime it deliberately avoids.

System-level implication

Modern VLA deployment is also an execution problem. LeRobot’s asynchronous inference decouples action execution from prediction, and its Real-Time Chunking documentation addresses pauses and discontinuities when chunks arrive late. FlowEdge’s cooperative steps are the low-level mechanism; the proposed FlowEdge Relay is where freshness, overlap, backpressure, and distributed scheduling belong.