Every model you can run.

Every job, to its own engine.

A PC already has five compute engines that were never introduced to each other. The NYMPH firmware routes every operation — language, vision, voice, retrieval, draft — to the optimal engine, and frees or recruits your GPU as needed.

Commercial rule: parallelism is guaranteed (each chip runs its own thing). Splitting one model across 2 chips (sharding) is roadmap — Pulsar2 allows it, pending on-silicon validation. Never promise "one 48-TOPS brain" if the SDK delivers "two 24-TOPS brains."

Your PC, as is, running local AI.

The honest floor: this is community-measured, free and local, without buying anything. Two machine profiles.

The golden rule: in dense, the usable ceiling is what fits in VRAM — ~14B on the Gamer, ~24B on the Creator; crossing it drops you to 1–5 tok/s. In MoE the active count rules, not the total: a 35B-A3B runs faster than a 14B dense. "80B at 30 tok/s on consumer" does not exist — a usable 80B gives ~15.

Image is compute-bound: the Gamer runs almost the same models, just slower; VRAM only blocks the biggest ones. Standard tool: ComfyUI.

Local video is minutes-per-clip except LTX — for iteration and B-roll, not clip-after-clip like the cloud. The 8 GB Gamer is essentially LTX-only.

The card doesn't compete on speed. It adds what's missing.

2× Axera AX8850 · 48 TOPS NPU (+6 orchestration) · 16 GB of AI memory · 64 GB persistent · ~16 W. The firmware already compiles and passes smoke 9/9 in simulation. Zero tok/s measured on Axera — every speed here is ceiling/projection until physical silicon.

The hard read: large dense = capacity, not speed (26B ~1.5–2.5 tok/s, for background agents). The truly fast = MoE + Mamba + spec-decode — that's where the projected double digits are. The defensible headline: "runs 26–30B MoE no gaming GPU can load, at ~16 W, always-on; targets ~9–11 tok/s — pending on-silicon measurement."

The same PC, before and after NYMPH.

NYMPH doesn't change what your GPU already does — it keeps it and adds what the GPU can't. That's why "with NYMPH" retains everything from "without NYMPH" and frees the GPU on top.

The sentence that sums up both: the PC alone is already fast — NYMPH doesn't compete there. It adds what no GPU gives itself: persistence, always-on with a free GPU, an extra brain above your VRAM ceiling, and auditable isolation. Like in the 90s your PC had no sound → Sound Blaster; in 2026 it can't run AI without sacrificing everything else → NYMPH.

Where each weight lives.

Decode is memory-bound: speed is set by the bandwidth of the tier where the active weight lives. The pooled hybrid adds tiers up to ~50–100 GB (roadmap).

Host↔card bottleneck: PCIe Gen2 (~0.9 GB/s effective). That's why expert-streaming and the pooled hybrid live or die on cache hit-rate, not raw bandwidth.

PC alone vs PC + NYMPH AX1 Premium.

The PC alone is already good at speed; NYMPH doesn't compete there. It adds the four things the PC can't give itself: persistence, always-on, extra model capacity, and auditable isolation. vs buying a $4,000 AI machine (DGX Spark): that wins raw tok/s and model ceiling, but it's another machine; AX1 empowers the one you already own and keeps your GPU free.

What governs every number.

Method: "without NYMPH" figures = community measurements/aggregates (llama.cpp, Unsloth, public benchmarks) or roofline calibrated against them. "With NYMPH" figures = bandwidth ceiling (roofline over spec) or own simulation projection — no tok/s is measured on a physical AX-M1; what compiles and how much it occupies IS measured in the Pulsar2 compiler. All are ±30% bands and vary with host, context, quantization and cooling. The nymph-firmware compiles and passes smoke 9/9 in simulation; silicon backends are 1:1 drop-ins when the board ships. Cited patents = USPTO provisionals (require conversion to non-provisional/PCT). NYMPH is a trademark of Punky Tiger Labs, Inc. AX8850 and AX-M1 are products of Axera Semiconductor; AICore AX-M1 is by Radxa. © 2026 Punky Tiger Labs, Inc.