SYMATICS
The landscape · as of October 2026
Model types, and what each means for state-resident compute
| Model type | Example models | What sets it apart | Compute profile — MPU view |
|---|---|---|---|
| Today’s baselines | |||
| Autoregressive Transformer LLM | GPT, Claude, Gemini, Llama, DeepSeek, Qwen | Predicts the next token with self-attention over the full context; the dominant paradigm since 2017. | Prefill is compute-bound; decode is memory-bandwidth-bound. KV-cache grows with context. |
| Reasoning LLM (test-time compute) | OpenAI o-series, DeepSeek-R1, extended thinking | RL-trained to think longer before answering; trades more inference tokens for accuracy on math, code and logic. | Very decode-heavy; long outputs and a large KV-cache; bandwidth-bound. |
| Beyond LLMs — and what changes | |||
| Decision model | Jev (TypeSafe AI) | State plus typed questions in; a choice, score or calibrated probability out — no text generation. | Single-pass, short-context, high-frequency; prefill-like and compute-bound, no decode loop. |
| Neural operator (physics foundation model) | Accelerated Understanding | Learns operators that map functions to functions; predicts physical phenomena across 3D space and time. | FFT / spectral convolutions over huge grids; memory-capacity and bandwidth heavy. |
| Diffusion LLM | Mercury (Inception Labs) | Denoises a whole block of masked tokens in a few parallel passes instead of generating one token at a time. | Parallel denoising; compute-bound and systolic-friendly, with far fewer serial steps. |
| Energy-based reasoning model | Kona (Logical Intelligence) | Learns an energy function over reasoning traces and finds solutions by minimizing it — full traces at once in latent space. | Iterative latent-space optimization; compute-bound, small models. |
| JEPA world model | V-JEPA 2, LeWorldModel (AMI Labs) | Predicts abstract representations of video and sensor data rather than pixels or tokens — for physical understanding and planning. | Video encoders plus predictors; vision-style compute with moderate memory. |
| Spatial / 3D world model | Marble (World Labs) | Generates persistent, explorable 3D worlds from text or images — spatial intelligence rather than language. | 3D generation and rendering; compute- and memory-heavy. |
| Liquid foundation model | LFM, LFM2.5 (Liquid AI) | Non-Transformer models derived from continuous-time neural networks; efficient on-device inference with a small memory footprint. | Edge-optimized; low bandwidth with recurrent-style state. |
| State-space model (Mamba) | Mamba, Cartesia Sonic | Linear-time recurrent sequence model with selective state and no attention; constant memory per token. | Scan / recurrence; bandwidth-light, with no KV-cache growth. |
| Hierarchical / recursive reasoning | HRM (Sapient Intelligence), TRM (Samsung SAIL) | Tiny networks that loop internally over latent state to refine answers — tens of millions of parameters. | Tiny weights, many iterations; on-chip-SRAM friendly. |
Third-party models are independent examples for context — not endorsements, partnerships or benchmarks. Profiles are architectural characterizations, not measured results. This page is maintained as new model paradigms emerge.
Why this matters
Different paradigms. One constant: state.
Every paradigm above keeps some form of state at its center — KV-caches, latent traces, recurrent state, world representations. The MPU architecture is investigated around exactly this: state that stays resident, structured, and continuously available, whatever form the model takes.