◆ hubSovEValSwarmEngineMAPM2026-10-17T01:43:22Z style="color:#8b949e;float:right">

SwarmEngine — swarm formation metrics

generated 2026-10-17T01:43:23+00:00 · schema formation-cells-v1 · auto-refresh 5min · spec: factory/capacity/SWARMENGINE-v0-SPEC.md

LIVE: GPU0: 524/16311 MiB (3% util) · GPU1: 14622/16311 MiB (0% util) · RAM available: 14.6 GiB · load1: 6.50 · :8010 DOWN · :11434 up (non-200 /health) · :11435 up (non-200 /health) · :11436 up (non-200 /health)

Width knees (knee not peak)

model	knee	agg peak t/s	useful width	note
ornith:9b	32	2241.8	32
qwen3.8-27b:q3xl	32	701.2	32
kat-coder-v2.5-dev:q4_K_M	32	2903.4	32
qwen3-coder-next:latest	32	2599.9	32
ornith-1.5:35b	32	1052.6	32
qwen3-next:80b	32	1143.9	32
gpt-oss:20b	32	167.2	24
devstral-small-2:24b	-	-	32
gemma4:12b	-	-	24
qwen3.5-122B:q3xl	16	50.9	16
glm-air:ud-q3xl	192	1320.7	64	GPU-PROMO

Formation cells (receipted)

cell	what	verdict
`hier_glm_orarith`	HIER: GLM-Air orchestrator -> ornith:9b workers	E2E 11.3-15.5s at N=2-8; ~4-5 HIER cycles/min production cadence
`debate_glm_ornith_bolt`	DEBATE/CONSENSUS: GLM-Air + ornith + Bolt blind; GLM-Air arbitrates	unanimous 12/12 correct incl. arbiter on three classic trap questions
`contention_hier_under_load`	CONTENTION: HIER(N=4) while 32-wide ornith background load saturates lane A	HIER loop E2E is saturation-resilient: 1.08x under full 32-wide background load
`sequential_chain`	SEQUENTIAL: ornith drafts -> Bolt refines -> GLM finalizes (context handoff)	Bolt is the binding leg in every formation it joins (43-59s) — deep-thinker role only, never on the hot path
`vram_tunnel`	VRAM TUNNEL: lane A swap ornith -> gpt-oss:20b -> ornith (bench-scope, operator tunnel-tes	through-swap ~30s for 13GB model (load-bandwidth-bound); back-swap 5.9s (warm cache) — production-viable transition
`glm_gpu_attention_promotion`	OPTIMIZATION: GLM-Air CPU -> GPU-attention (all experts CPU, attn+KV GPU1)	lib-split law env recipe (GGML_BACKEND_PATH=cuda_v13 + LD_LIBRARY_PATH) proven on bundled llama-server
`all_lanes_hot`	ALL-LANES-HOT: multi-model concurrent map — finding the true ceiling
`hier_1792194619`	hier
`trace_replay_1792194954`	trace_replay
`trace_replay_1792194978`	trace_replay
`cpu_seating_probes`	CPU-SEATING: active-params physics probe (lane B, num_gpu:0 forced)
`lightning_lane_state`	LIGHTNING (qwen3.5-122B:q3xl) receipts + self-eviction
`supervl_truncation_discovery`	SUPER VL download truncation (handoff claim was FALSE)
`bolt_restore`	BOLT restore per own-libggml law
`knee_ollama_//127.0.0.1_11435/qwen3.5-122B_q3xl`
`hier2_1792196756`	hier2
`hot_1792196798`	hot
`contention_1792196811`	contention
`blackboard_1792197193`	blackboard
`prefix_cache_ab`	A/B: --cache-reuse 2048 vs default (agent-style shared-prefix workload, 4x1077-tok prompts	MARGINAL - 16% cold-prefill win, warm identical. llama-server default slot cache ALREADY reuses exact repeated prefixes (1102 tok/s warm). Prefix leve
`chain_1792198400`	chain
`a1_tool_router_chain`	A1 TOOL-ROUTER CHAIN: GLM(brain) -> A1(tool-router, CPU lane) -> ornith(executor)	A1 viable but LATENCY-BOUND: its leg is 38s (thinking + tool discipline at 14 t/s CPU) = 3-4x a Glimmer sub-orch (8-16s class). Use A1 where structure
`blackboard_1792198472`	blackboard
`trace_replay_1792198490`	trace_replay
`taic_1792199196`

Roster

disposition	models
KEEP	glm-air:ud-q3xl, muse-glimmer:30b-q3xl, qwen3.8-27b:q3xl, gpt-oss:20b, qwen3.5-35b-a3b:iq4_xs, qwen3.5-122B:q3xl, nemotron-3.5-bolt:74b-iq2xxs, ornith:9b, ornith-1.5:35b, kat-coder-v2.5-dev:q4_K_M, qwen3-coder-next:lates
ARCHIVE	qwen3-next:80b
TOOLING_POOL	class, models, note
DROP-DECIDED

Seating laws

CPU seat: active params per token decide speed (MoE A3B 17-19 t/s, dense-30B 5.4 t/s). Density is punished by CPU.
RAM ceiling: ONE CPU-seated 100B-class brain + ONE A3B CPU lane + 2 GPU lanes = the 60GB frontier. More = OOM.
GPU lanes are insulated: CPU-side contention never touched ornith (60 t/s) or Glimmer (11.4 t/s) under all-hot load.
CUDA_VISIBLE_DEVICES='' does NOT hide GPUs from ollama — force num_gpu:0 per request for CPU lanes.
num_thread caps: CPU lanes only (6). Capping GPU-lane threads starves them (ornith 60->20 t/s).
Knee not peak: sustainable width where n/n succeed; legacy peak claims died under this rule (2,241 -> 61.8 real).
Bolt-class deep thinkers bind every formation (43-59s). Never on the hot path.
At 4+ model occupancy restarts cost minutes (page-faulting). Keep models resident.