The mesh-brain end-state the near-term deployment-shape work builds toward — tracked here so the near-term PR can stay small.
End-state: one lobe per box
| Box |
Lobe it hosts |
Notes |
| DGX Spark GB10 (128 GB) |
Qwen cortex only (sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP) |
Reclaims the co-residency tax — candidate to return from 128K to the 256K-native solo shape |
| Jetson AGX Thor 128 GB (sm_110) |
Gemma senses only (Gemma 4 12B multimodal) |
Full-context senses instead of the 32K co-resident trim |
| Jetson AGX Orin 64 GB |
Small-model lobes (minor-class generate + 0.6B pooling gears) |
Unvalidated card — per the #108 rule it ships nothing until someone boots one |
Where the embedder/reranker/audio gears land in the end-state is part of this issue's design work (they are cheap enough to co-reside almost anywhere).
Path
- Near-term (spec being cut now, devague frame in progress): deployment shape becomes a first-class, safely-selectable axis —
machine-as-brain (default: one box hosts every lobe it can — not everyone has many machines) vs per-box mesh-brain lobe profiles. First mesh-brain move: Spark drops Gemma senses (keeps cortex + embedder + reranker + audio), Thor drops the Qwen cortex (keeps senses + the rest).
- This issue (future PR): specialize all the way down to one lobe per box, per the table above.
Constraints carried forward
The mesh-brain end-state the near-term deployment-shape work builds toward — tracked here so the near-term PR can stay small.
End-state: one lobe per box
sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP)Where the embedder/reranker/audio gears land in the end-state is part of this issue's design work (they are cheap enough to co-reside almost anywhere).
Path
machine-as-brain(default: one box hosts every lobe it can — not everyone has many machines) vs per-boxmesh-brainlobe profiles. First mesh-brain move: Spark drops Gemma senses (keeps cortex + embedder + reranker + audio), Thor drops the Qwen cortex (keeps senses + the rest).Constraints carried forward
Substrate: spec+plan: lobes fits the machine it lands on — per-machine hardware profiles #108's per-machine profiles (feasible roles + model-per-role + traits + goldens) — composition is data over that schema, not a new code fork. verify on the GB10 before t3 ships: is the fp8-KV crash checkpoint-driven, and is VLLM_ATTENTION_BACKEND truly dead on the pinned image? #109 verified the trait layer on the GB10 (fp8-KV is an sm_110-conditional trait; the attention-backend env is dead on the shared pinned nightly).
Honesty (Advertised per-role endpoint (:8000) 404s while ready=true — clients dialing it degrade instantly (lobes-cli#87 shape regressed) #92): a box that doesn't host a lobe must not advertise it — capabilities omit it and the gateway 4xxes it, never silently reroutes.
Reclamation: dropping a lobe must actually free its budget for the remaining lobes (util/context re-derived), or specialization buys nothing.
One-box users are unaffected: machine-as-brain stays the default; the end-state is opt-in per box.
Depends on the cross-box question the near-term spec must answer first: does a mesh-brain box's gateway proxy absent roles to the peer that hosts them, or do consumers address each box directly?
lobes (Claude)