Kvasir blog

Expert-Sharded Swarm Inference: The Design — 86% of a 122B MoE is 12,544 independent 5.3 MB experts. Slice the model at that grain and a phone can carry a real share of frontier inference.
From Blueprint to Hardware: What's Verified, and the Keystone — A recap of the verification campaign — design through M2-core proved on a real 122B — and the one integration piece that unlocks the rest.
Numerical Equivalence Across Heterogeneous Backends — CUDA, ROCm, Adreno and CPUs will never agree bit-for-bit. That the swarm still produces one coherent model is a designed property, not luck.
NVIDIA Blackwell Joined the Swarm — A GB10 Grace Blackwell computed real 122B expert-FFN slices in CUDA and matched AMD ROCm bit-for-bit (cosine 1.0000000000) and Grace ARM CPU within tolerance. The cross-backend matrix is complete.
Hardening the Money Path: Transaction Security in KVR Settlement — Three real vulnerability classes — payment-signature replay, unauthenticated reward minting, and race double-spends — found, exploited in tests, and closed in the gateway's settlement service.
Phase 0 — The Engine: linkcpp, a Control Plane for inference engine — inference engine ships a capable RPC data plane but no control plane. linkcpp adds the missing half — discovery, planning, launch and gateways — around stock binaries.
Phase 1 — The Ring: Pipeline Inference Without a Master — Every device loads only its layer window and passes a small hidden-state boundary to its neighbor. No node holds the model; no central master exists.
Phase 2 — Inside a 122B MoE: Why the Weights Want to Be Sharded — A tensor-level analysis of Qwen3.5-122B: 86% of the bytes are 12,544 independent expert slabs, each one a clean byte-range copy away from standing alone.
Phase 3 — Backbone Expert RAM Offload (M0) — Stream MoE expert FFNs from CPU RAM instead of VRAM, and a single 64 GB coordinator holds a 122B — with no graph surgery.
Phase 4 — The Expert-Slice Data Path (M1) — A weak device downloads a few 6 MB experts, not a 1.4 GB layer — and sharded compute matches monolithic to 3.6e-12.
Phase 5 — Distributed Expert Dispatch (M2) — A live 122B decode hands one layer's expert compute to a separate worker process over TCP — and predicts exactly the same token.
Phase 6 — The Expert Coverage Market (M3) — Weak nodes see which (layer, expert-range) is scarcest and highest-reward, and fill it themselves — the proven layer market, regrained.
Phase 7 — Batched Dispatch Throughput (M4) — The swarm is a throughput fabric, not a latency play: batching dispatch calls amortizes per-request overhead 77× per token.
A Phone Joined 122B Inference — A Galaxy S25 autonomously downloaded its expert slice from the hub and computed one layer's experts every step of a live 122B decode. The output was correct.
The Kvasir Economy: A Virtuous Cycle of Cost and Reward — A decentralized inference network only works if the price consumers pay and the reward nodes earn reinforce each other. Here is the flywheel we're building toward, the spirals that kill it, and the three invariants that keep it turning.
A GPU Across the Internet Joined 122B Inference — A Blackwell workstation in another city dialed one outbound 443 connection and computed experts for a live 122B decode — byte-identical to a local run, and paid in KVR for the work it did.
Making WAN Dispatch Cheap: A Grounded Roadmap — Remote expert dispatch works and is byte-identical — but a WAN decode is round-trip-bound. Here's the plan to cut the cost, grounded in production numbers from DeepSeek, Petals and others (roadmap, not shipped).
The 200G Question — Our swarm hubs can already link at 200 Gb/s with parts on the shelf — one is built into the GB10. Here is what a fat pipe buys a distributed MoE, and the one thing it can't.
The swarm that grows under load — A giant model that borrows help only when it needs it — the MoE swarm now scales itself to traffic: tight and fast when quiet, wide and parallel when busy.