# Cross-Model Inference Pareto Study on the Intel Arc Pro B70

**Models:** AgentWorld-35B-A3B · NEX2 (Nex-N2-mini) · SIQ-1-35B · DREAM-v0-Instruct-7B
**Hardware:** 1× Intel Arc Pro **B70** — 256 Xe cores, 30.3 GiB, 230 W cap, Level-Zero/SYCL.
**Engine:** `llama.cpp` SYCL "moe-ready" build. **Date:** 2026-06-27.

## 1. Models

| Model | Arch | Params | Quant | Size | Decode paradigm |
|---|---|---|---|---:|---|
| AgentWorld-35B-A3B | `qwen35moe` (40L, 256 exp/8 act, GQA 16/2) | 34.7B / ~3B active | Q5_K_M | 24.7 GB | autoregressive |
| NEX2 (Nex-N2-mini) | `qwen35moe` (identical topology) | 34.7B / ~3B active | Q5_K_M | 24.7 GB | autoregressive |
| SIQ-1-35B | `qwen35moe` (identical topology) | 34.7B / ~3B active | Q5_K_M | 24.7 GB | autoregressive |
| Ornith-1.0-35B | `qwen35moe` (identical topology) | 34.7B / ~3B active | Q5_K_M | 24.7 GB | autoregressive |
| DREAM-v0-Instruct-7B | `dream` (28L, dense, 28/4 heads) | 7B dense | Q4_K_M | 4.7 GB | **diffusion** |

The four 35B models are independent fine-tunes of the **same Qwen3.5-35B-A3B base** — a controlled
"same-architecture, different-weights" set. DREAM is an architectural outlier (masked **diffusion** LM).

## 2. Method

- **Autoregressive Pareto** (`bench/concurrency_pareto_guarded.sh`, `bench/context_pareto.sh`):
  `llama-batched-bench`, ship config `-ngl 99 -fa on -ctk/-ctv q8_0 -b 8192 -ub 4096`,
  `GGML_SYCL_DISABLE_DNN=1`. Crash-safe (each cell a separate fsynced process) + per-cell timeout (a cell
  exceeding 300 s = memory-spill thrash → stop). Concurrency: realistic agent turn (2048 prompt + 256 gen),
  sweep agents. Context: sweep per-agent depth × max-agents-that-fit.
- **Diffusion Pareto** (`bench/dream_diffusion_pareto.sh`): `llama-diffusion-cli`, sweep `--diffusion-steps`
  at fixed `n_predict=128` (confidence-based algorithm, timestep schedule, `eps=0.001`). No `-np`
  concurrency exists for a single-stream diffusion CLI; the lever is the step count.

## 3. Concurrency Pareto (aggregate decode tok/s vs agents)

| agents | AgentWorld | NEX2 | SIQ | Ornith |
|---:|---:|---:|---:|---:|
| 1 | 76.4 | 75.8 | 75.5 | 75.9 |
| 8 | 107.8 | 103.3 | 106.5 | 106.8 |
| 16 | 127.8 | 121.7 | 126.0 | 126.3 |
| 32 | 159.9 | 151.5 | 157.6 | 157.9 |
| **48 (peak)** | **178.3** | **168.9** | **175.5** | **175.5** |
| 56 | **cliff** | **cliff** | **cliff** | **cliff** |

**Finding:** the four same-architecture fine-tunes **collapse onto one curve within ~5%** (peak spread
168.9–178.3, three of four at 175–178), with an **identical hard memory cliff at 56 agents** (aggregate
collapses ~18× — the 30 GiB wall). Single-agent throughput is near-identical (75.5–76.4); the small spread
at high concurrency is
run-to-run + thermal variance (hours of continuous load), not an architectural difference. **Throughput is
architecture-determined; these models are interchangeable on speed and differ only in quality.** Knee = 32
agents (≈90% of peak, safe headroom); never serve ≥56.

## 4. Context Pareto (max aggregate decode vs per-agent context depth)

| per-agent depth | max agents | AgentWorld | NEX2 | SIQ | Ornith |
|---:|---:|---:|---:|---:|---:|
| 512 | 48 | 209.0 | 201.2 | 206.4 | 206.7 |
| 2,048 | 48 | 177.7 | 169.5 | 175.6 | 175.9 |
| 8,192 | 28 | 105.2 | 101.1 | 104.8 | 104.8 |
| 16,384 | 15 | 65.2 | 63.4 | 65.2 | 65.2 |
| 32,768 | 6 | 34.4 | 33.9 | 34.4 | 34.4 |
| 65,536 | 3 | 22.4 | 22.2 | 22.4 | 22.3 |

**Finding:** again the four overlay (within ~4%). Concurrency and per-agent context trade for the same
30 GiB (`agents × depth ≤ ~262 k KV-tokens` AND `agents ≤ ~48` slots). Aggregate throughput falls 209 → 22
tok/s as per-agent context climbs 512 → 65,536; shallow depths bind on the 48-slot wall, deep depths on the
262 k-token KV ceiling.

## 5. DREAM-7B diffusion Pareto (a different axis entirely)

| diffusion steps | tok/s | ms/step | quality (spot-check) |
|---:|---:|---:|---|
| 8 | **45.9** | 348 | incoherent ("the the the the") |
| 16 | 26.3 | 305 | broken |
| 32 | 13.9 | 287 | partial |
| 64 | 7.3 | 275 | partial |
| 128 | 3.7 | 273 | terse |
| 256 | **1.84** | 271 | **coherent** ("scattering of sunlight by the atmosphere…") |

**Finding:** diffusion has **no concurrency or context axis** — its Pareto is a **speed↔quality dial** on the
step count. Throughput is `n_predict / (steps × ~275 ms)`; quality rises monotonically with steps. The
usable-quality point (~256 steps) is **1.8 tok/s** — ~40× slower than the 35B-A3B MoE models' 75 tok/s
single-stream (note DREAM is a 7B *dense* model, so per-step it's cheap, but it pays ~256 full passes for one
coherent block). On the B70, masked-diffusion is **not** competitive with the MoE-A3B autoregressive models
for throughput-at-quality; its appeal (parallel token refinement, controllable compute) doesn't convert to a
B70 throughput win at this size/quant.

## 6. Cross-paradigm summary

| | 35B-A3B MoE (AR) | 7B diffusion (DREAM) |
|---|---|---|
| Best single-stream tok/s | ~75 | 45.9 (8 steps, incoherent) / 1.8 (256, coherent) |
| Fleet aggregate (peak) | 169–178 (48 agents) | n/a (single-stream) |
| Scaling lever | concurrency (np) + context | diffusion steps |
| Quality dial | sampling/quant | step count (explicit) |
| B70 verdict | throughput champion; serve `-np 32` | speed/quality-dialable but slow at quality |

## 7. Agent-collaboration findings (companion experiments, AgentWorld)

- **Swarm vs solo (separable breadth):** 48-agent independent fan-out beats solo — runbook coverage 21→39/40,
  edge-cases 107→851. (`SWARM_VS_SOLO_B70.md`)
- **Collaborative game build (coupled correctness):** harness *architecture* dominates. A component-
  decompose CrewAI harness produced bigger but **broken** games (unwinnable Frogger, unshootable maze, code
  quality 27/50); the canonical whole-game-worker + merge + regression-guard teamwork harness produced
  **correct** games (42/41 vs solo 40/38). (`crew_games/eval_games/CODE_QUALITY_EVAL.md`,
  `crew_games/teamwork_out/TEAMWORK_RETEST_B70.md`)

## 8. Reproducibility

Harnesses: `bench/{concurrency_pareto_guarded,context_pareto,dream_diffusion_pareto}.sh`,
`bench/{serve_np,concurrent_agents}.py`, `crew_games/crew_build.py`, `agentic-arcade/teamwork/`.
Raw data: `results/{,nex2_,siq_}{concurrency,context}_pareto*/`, `results/dream_diffusion_pareto/`.
Charts: `results/charts/*.png` (per-model + overlays + DREAM). All runs single-GPU B70, ship config above.