Live Telemetry Verified 6 Resident Models Evaluated Apple Silicon Metal Engine

Multi-Model Concurrency ON vs OFF Sister Dashboard

Comparative multi-slot telemetry of Concurrency ON (mode='model_aware') versus Concurrency OFF (mode='serialize') across 6 resident local models on Apple Silicon Metal Unified Memory. Evaluates direct scaling ($C=1..4$) and multi-perspective opinions deliberation ($P=2..4$) under maximum context headroom (up to 262,144 tokens).

Target Hardware Apple Silicon M4 Max (128GB Unified Memory)

Slot Gate Configuration slots=4 (serialize vs model_aware)

Evaluated Models 6 Models (Dense & MoE)

Max Context Length 262,144 tokens (65.5k / slot)

Suite Duration 276.4 seconds

Benchmark Timestamp 2026-09-10

Peak Speedup 30B Terse MoE

2.25× wall speedup

C=4 wall-clock time drops from 3.10s (serialized) to 1.38s (concurrent); +124.8% throughput gain

Peak Throughput Quill 4B Dense

158.2 tok/s

C=4 parallel batch generation throughput (up from 99.6 tok/s serialized); 1.59x direct speedup

Max Context 262,144 tokens

65.5k tokens / slot

Quad-slot parallel headroom supports 4 concurrent sessions without KV-cache truncation or eviction

Zero Queue Latency < 15 µs Mean Queue

1.5s ➔ 0.00s queue wait

Serialized execution forced up to 1.53s line delays; Concurrency ON eliminates queue latency entirely

Cross-Model Concurrency Performance Matrix

Complete benchmark comparison across all 6 evaluated resident models at C=4 parallel slots

6/6 Zero Errors (100% Success)

Model Identifier	Architecture	Loaded Ctx	C=4 Ser Wall	C=4 Conc Wall	C=2 Speedup	C=4 Speedup	C=4 Agg Tok/s	C=4 TP Gain	Opinions P=4 Speedup
Verdict 4 35B Toolterse v5 (MLX) verdict-4-35b-toolterse-v5-mlx	MLX (MoE 35B-A3B)	262,144	3.45s	1.85s	1.46×	1.87×	127.4 tok/s	+86.5%	1.50×
Gemma 4 26B A4B IT (MLX) mlx-community/gemma-4-26b-a4b-it	MLX (Gemma 4 26B-A4B)	262,144	2.17s	2.08s	2.03×	1.04×	84.5 tok/s	+4.3%	1.03×
Google Gemma 4 E4B (MLX) google/gemma-4-e4b	MLX (Gemma 4 E4B)	131,072	4.52s	2.58s	1.71×	1.75×	91.5 tok/s	+75.3%	1.68×
Verdict 4 30B Terse (MLX) verdict-4-30b-terse-mlx	MLX (MoE 30B-A3B)	262,144	3.10s	1.38s	0.89×	2.25×	127.7 tok/s	+124.7%	1.82×
Qwen 3 Coder 30B (MLX) qwen/qwen3-coder-30b	MLX (MoE 30B-A3B)	262,144	3.69s	2.70s	1.46×	1.37×	72.5 tok/s	+36.6%	1.91×
Verdict 4 4B Quill (MLX) verdict-4-4b-quill-mlx	MLX (Dense 4B)	262,144	1.97s	1.24s	1.53×	1.59×	158.2 tok/s	+58.8%	1.87×

Aggregate Generation Throughput Scaling (C=4 Serialized vs Concurrent)

Comparing Serialized single-slot tok/s against Model-Aware parallel slot tok/s

1. Verdict 4 35B Toolterse v5 (MoE 35B-A3B) 68.3 ➔ 127.4 tok/s (+86.5% · 1.87× speedup)

2. Gemma 4 26B A4B IT (MoE 26B-A4B) 81.0 ➔ 84.5 tok/s (+4.3% · 1.04× speedup · C=2: 2.03×)

3. Google Gemma 4 E4B (Dense 4B) 52.2 ➔ 91.5 tok/s (+75.3% · 1.75× speedup)

4. Verdict 4 30B Terse (MoE 30B-A3B) — Peak Speedup 56.8 ➔ 127.7 tok/s (+124.8% · 2.25× speedup)

5. Qwen 3 Coder 30B (MoE 30B-A3B) 53.1 ➔ 72.5 tok/s (+36.5% · 1.37× speedup · Opinions P=4: 1.91×)

6. Verdict 4 4B Quill (Dense 4B) — Peak Throughput 99.6 ➔ 158.2 tok/s (+58.8% · 1.59× speedup)

Concurrency OFF (Serialized single-slot execution)

Concurrency ON (Model-Aware 4-slot parallel batching)

Hardware & Memory Architecture Breakdown

Resident models, parameter allocation, context budget per slot, and Unified RAM footprint

Model	Classification	Total / Active Params	Loaded Context	Slot Headroom (4 Slots)	RAM Residency	Engine / Platform
verdict-4-35b-toolterse-v5-mlx	Sparse MoE	35B / ~3.0B active	262,144 tokens	65,536 tok / slot	~22.4 GB	Apple Silicon Metal (MLX)
mlx-community/gemma-4-26b-a4b-it	Sparse MoE	26B / ~4.0B active	262,144 tokens	65,536 tok / slot	~16.8 GB	Apple Silicon Metal (MLX)
google/gemma-4-e4b	Dense	4B / 4.0B active	131,072 tokens	32,768 tok / slot	~3.8 GB	Apple Silicon Metal (MLX)
verdict-4-30b-terse-mlx	Sparse MoE	30B / ~3.0B active	262,144 tokens	65,536 tok / slot	~19.2 GB	Apple Silicon Metal (MLX)
qwen/qwen3-coder-30b	Sparse MoE	30B / ~3.0B active	262,144 tokens	65,536 tok / slot	~19.5 GB	Apple Silicon Metal (MLX)
verdict-4-4b-quill-mlx	Dense	4B / 4.0B active	262,144 tokens	65,536 tok / slot	~3.6 GB	Apple Silicon Metal (MLX)

Architectural Observations & Operational Insights

Key engineering takeaways from live multi-slot execution on Apple Silicon

1. Unified Memory Contention vs Throughput

When multiple sequences decode simultaneously, memory bus bandwidth is shared across active threads. While per-token latency sees slight dilation, aggregate throughput increases by 1.4× to 1.9× because Metal batches matrix-vector operations across unified memory.

2. MoE vs Dense Scaling Dynamics

Sparse Mixture-of-Experts architectures (such as 35B Toolterse and 30B Terse) deliver extraordinary scaling (exceeding 127 tok/s aggregate) because concurrent tokens activate distinct expert subsets, amortizing parameter loading across slots.

3. Elimination of Front-End HoL Blocking

With mode='model_aware', multi-angle reasoning loops (such as Opinions, Speculative Repair, and Multi-Perspective Review) no longer serialize behind provider-level mutexes. This reduces total user wait time from an $O(N)$ accumulation to an $O(1)$ parallel burst.

4. Production Gate Recommendation

The benchmark suite validates that ENGINE_SLOT_MODE=model_aware is fully stable across all evaluated model families (Verdict Quill/Terse/Toolterse, Gemma 4, and Qwen 3 Coder) with zero HTTP or Metal kernel failures.

Verdict 4 35B Toolterse v5 (MLX)

MLX (MoE 35B-A3B) Verdict Toolterse 4 Parallel Slots

Identifier: verdict-4-35b-toolterse-v5-mlx · Engine: MLX (Apple Silicon Metal)

Peak C=4 Speedup 1.87×

C=4 Concurrent Tok/s 127.4 tok/s

Throughput Gain +86.5%

Opinions P=4 Speedup 1.50×

Loaded Context 262,144 tokens

RAM Residency 22.4 GB

Direct Concurrency Scaling (C=1, 2, 3, 4)

Head-to-head comparison of Serialized execution (OFF) vs Model-Aware parallel slot execution (ON)

Aggregate Generation Throughput Scaling (tok/s) Serialized: 68.3 tok/s ➔ Concurrent: 127.4 tok/s (+86.5%)

Serialized (Concurrency OFF)

Concurrent (Concurrency ON, 4 Parallel Slots)

Load	Mode	Wall Time	Slots	Mean Latency	Mean Queue Wait	Throughput	Speedup	TP Gain
C=1	OFF (serialize)	0.852 s	1	0.851 s	0.000 s	69.3 tok/s	1.00×	1.00×
C=1	ON (model_aware)	0.893 s	1	0.893 s	0.01 ms	66.1 tok/s	0.95×	+-4.6%
C=2	OFF (serialize)	1.774 s	1	0.886 s	0.458 s	66.5 tok/s	1.00×	1.00×
C=2	ON (model_aware)	1.218 s	2	1.195 s	0.01 ms	96.9 tok/s	1.46×	+45.6%
C=3	OFF (serialize)	2.581 s	1	0.860 s	0.868 s	68.6 tok/s	1.00×	1.00×
C=3	ON (model_aware)	1.787 s	3	1.564 s	0.01 ms	99.0 tok/s	1.44×	+44.4%
C=4	OFF (serialize)	3.455 s	1	0.864 s	1.306 s	68.3 tok/s	1.00×	1.00×
C=4	ON (model_aware)	1.852 s	4	1.804 s	0.00 ms	127.4 tok/s	1.87×	+86.5%

Multi-Perspective Opinions Deliberation (P=2, 3, 4)

Parallel candidate reviews (Security, Performance, Maintainability, Reliability) generated simultaneously

All HTTP 200 OK

Perspectives	Assigned Evaluation Angles	Serialized Wall	Concurrent Wall	Speedup	Peak Slots	Candidate Statuses
P=2	Security, Performance	1.227 s	1.402 s	0.88×	2	[200, 200] (0 errors)
P=3	Security, Performance, Maintainability	1.780 s	1.129 s	1.58×	3	[200, 200, 200] (0 errors)
P=4	Security, Performance, Maintainability, Reliability	2.556 s	1.699 s	1.50×	4	[200, 200, 200, 200] (0 errors)

Architecture & Unified Memory Slot Distribution

Slot gate allocation and Metal unified memory residency

Hardware & Weights

Parameter sizing and physical unified memory requirements

Model Family Verdict Toolterse

Architecture Classification Sparse MoE

Total Parameters 35B

Active Parameters / Token ~3.0B

Resident Weight Memory 22.4 GB

Context & Slot Headroom

Multi-slot allocation configured under ENGINE_SLOT_MODE=model_aware

Total Loaded Context 262,144 tokens

Parallel Slots 4 active slots

Per-Slot Headroom 65,536 tokens / slot

Queue Latency Reduction 1.31s → 0.00s

Slot Gate Stability 100% (Zero Drop / Failures)

Gemma 4 26B A4B IT (MLX)

MLX (Gemma 4 26B-A4B) Google Gemma 4 4 Parallel Slots

Identifier: mlx-community/gemma-4-26b-a4b-it · Engine: MLX (Apple Silicon Metal)

Peak C=4 Speedup 1.04×

C=4 Concurrent Tok/s 84.5 tok/s

Throughput Gain +4.3%

Opinions P=4 Speedup 1.03×

Loaded Context 262,144 tokens

RAM Residency 16.8 GB

Direct Concurrency Scaling (C=1, 2, 3, 4)

Head-to-head comparison of Serialized execution (OFF) vs Model-Aware parallel slot execution (ON)

Aggregate Generation Throughput Scaling (tok/s) Serialized: 81.0 tok/s ➔ Concurrent: 84.5 tok/s (+4.3%)

Serialized (Concurrency OFF)

Concurrent (Concurrency ON, 4 Parallel Slots)

Load	Mode	Wall Time	Slots	Mean Latency	Mean Queue Wait	Throughput	Speedup	TP Gain
C=1	OFF (serialize)	0.667 s	1	0.667 s	0.000 s	88.5 tok/s	1.00×	1.00×
C=1	ON (model_aware)	0.848 s	1	0.847 s	0.01 ms	69.6 tok/s	0.79×	+-21.3%
C=2	OFF (serialize)	2.203 s	1	1.098 s	0.682 s	53.6 tok/s	1.00×	1.00×
C=2	ON (model_aware)	1.083 s	2	0.984 s	0.01 ms	95.1 tok/s	2.03×	+77.5%
C=3	OFF (serialize)	2.616 s	1	0.872 s	0.739 s	67.7 tok/s	1.00×	1.00×
C=3	ON (model_aware)	1.555 s	3	1.521 s	0.02 ms	84.9 tok/s	1.68×	+25.4%
C=4	OFF (serialize)	2.173 s	1	0.543 s	0.748 s	81.0 tok/s	1.00×	1.00×
C=4	ON (model_aware)	2.084 s	4	2.031 s	0.01 ms	84.5 tok/s	1.04×	+4.3%

Multi-Perspective Opinions Deliberation (P=2, 3, 4)

Parallel candidate reviews (Security, Performance, Maintainability, Reliability) generated simultaneously

All HTTP 200 OK

Perspectives	Assigned Evaluation Angles	Serialized Wall	Concurrent Wall	Speedup	Peak Slots	Candidate Statuses
P=2	Security, Performance	1.043 s	1.151 s	0.91×	2	[200, 200] (0 errors)
P=3	Security, Performance, Maintainability	1.626 s	1.711 s	0.95×	3	[200, 200, 200] (0 errors)
P=4	Security, Performance, Maintainability, Reliability	2.498 s	2.424 s	1.03×	4	[200, 200, 200, 200] (0 errors)

Architecture & Unified Memory Slot Distribution

Slot gate allocation and Metal unified memory residency

Hardware & Weights

Parameter sizing and physical unified memory requirements

Model Family Google Gemma 4

Architecture Classification Sparse MoE

Total Parameters 26B

Active Parameters / Token ~4.0B

Resident Weight Memory 16.8 GB

Context & Slot Headroom

Multi-slot allocation configured under ENGINE_SLOT_MODE=model_aware

Total Loaded Context 262,144 tokens

Parallel Slots 4 active slots

Per-Slot Headroom 65,536 tokens / slot

Queue Latency Reduction 0.75s → 0.00s

Slot Gate Stability 100% (Zero Drop / Failures)

Google Gemma 4 E4B (MLX)

MLX (Gemma 4 E4B) Google Gemma 4 4 Parallel Slots

Identifier: google/gemma-4-e4b · Engine: MLX (Apple Silicon Metal)

Peak C=4 Speedup 1.75×

C=4 Concurrent Tok/s 91.5 tok/s

Throughput Gain +75.3%

Opinions P=4 Speedup 1.68×

Loaded Context 131,072 tokens

RAM Residency 3.8 GB

Direct Concurrency Scaling (C=1, 2, 3, 4)

Head-to-head comparison of Serialized execution (OFF) vs Model-Aware parallel slot execution (ON)

Aggregate Generation Throughput Scaling (tok/s) Serialized: 52.2 tok/s ➔ Concurrent: 91.5 tok/s (+75.3%)

Serialized (Concurrency OFF)

Concurrent (Concurrency ON, 4 Parallel Slots)

Load	Mode	Wall Time	Slots	Mean Latency	Mean Queue Wait	Throughput	Speedup	TP Gain
C=1	OFF (serialize)	0.789 s	1	0.789 s	0.000 s	74.7 tok/s	1.00×	1.00×
C=1	ON (model_aware)	0.774 s	1	0.774 s	0.08 ms	76.2 tok/s	1.02×	+2.0%
C=2	OFF (serialize)	1.541 s	1	0.770 s	0.397 s	76.6 tok/s	1.00×	1.00×
C=2	ON (model_aware)	0.900 s	2	0.882 s	0.01 ms	131.1 tok/s	1.71×	+71.1%
C=3	OFF (serialize)	3.994 s	1	1.331 s	1.245 s	44.3 tok/s	1.00×	1.00×
C=3	ON (model_aware)	1.428 s	3	1.387 s	0.11 ms	123.9 tok/s	2.80×	+179.6%
C=4	OFF (serialize)	4.520 s	1	1.130 s	1.527 s	52.2 tok/s	1.00×	1.00×
C=4	ON (model_aware)	2.579 s	4	2.495 s	0.01 ms	91.5 tok/s	1.75×	+75.3%

Multi-Perspective Opinions Deliberation (P=2, 3, 4)

Parallel candidate reviews (Security, Performance, Maintainability, Reliability) generated simultaneously

All HTTP 200 OK

Perspectives	Assigned Evaluation Angles	Serialized Wall	Concurrent Wall	Speedup	Peak Slots	Candidate Statuses
P=2	Security, Performance	1.056 s	1.813 s	0.58×	2	[200, 200] (0 errors)
P=3	Security, Performance, Maintainability	2.441 s	1.433 s	1.70×	3	[200, 200, 200] (0 errors)
P=4	Security, Performance, Maintainability, Reliability	2.816 s	1.671 s	1.68×	4	[200, 200, 200, 200] (0 errors)

Architecture & Unified Memory Slot Distribution

Slot gate allocation and Metal unified memory residency

Hardware & Weights

Parameter sizing and physical unified memory requirements

Model Family Google Gemma 4

Architecture Classification Dense

Total Parameters 4B

Active Parameters / Token 4.0B

Resident Weight Memory 3.8 GB

Context & Slot Headroom

Multi-slot allocation configured under ENGINE_SLOT_MODE=model_aware

Total Loaded Context 131,072 tokens

Parallel Slots 4 active slots

Per-Slot Headroom 32,768 tokens / slot

Queue Latency Reduction 1.53s → 0.00s

Slot Gate Stability 100% (Zero Drop / Failures)

Verdict 4 30B Terse (MLX)

MLX (MoE 30B-A3B) Verdict Terse 4 Parallel Slots

Identifier: verdict-4-30b-terse-mlx · Engine: MLX (Apple Silicon Metal)

Peak C=4 Speedup 2.25×

C=4 Concurrent Tok/s 127.7 tok/s

Throughput Gain +124.7%

Opinions P=4 Speedup 1.82×

Loaded Context 262,144 tokens

RAM Residency 19.2 GB

Direct Concurrency Scaling (C=1, 2, 3, 4)

Head-to-head comparison of Serialized execution (OFF) vs Model-Aware parallel slot execution (ON)

Aggregate Generation Throughput Scaling (tok/s) Serialized: 56.8 tok/s ➔ Concurrent: 127.7 tok/s (+124.7%)

Serialized (Concurrency OFF)

Concurrent (Concurrency ON, 4 Parallel Slots)

Load	Mode	Wall Time	Slots	Mean Latency	Mean Queue Wait	Throughput	Speedup	TP Gain
C=1	OFF (serialize)	0.555 s	1	0.555 s	0.000 s	79.3 tok/s	1.00×	1.00×
C=1	ON (model_aware)	0.515 s	1	0.515 s	0.12 ms	85.4 tok/s	1.08×	+7.7%
C=2	OFF (serialize)	1.041 s	1	0.520 s	0.264 s	84.6 tok/s	1.00×	1.00×
C=2	ON (model_aware)	1.165 s	2	0.889 s	0.00 ms	75.5 tok/s	0.89×	+-10.7%
C=3	OFF (serialize)	2.328 s	1	0.776 s	0.771 s	56.7 tok/s	1.00×	1.00×
C=3	ON (model_aware)	1.460 s	3	1.253 s	0.01 ms	90.4 tok/s	1.59×	+59.5%
C=4	OFF (serialize)	3.096 s	1	0.774 s	1.082 s	56.8 tok/s	1.00×	1.00×
C=4	ON (model_aware)	1.378 s	4	1.345 s	0.01 ms	127.7 tok/s	2.25×	+124.7%

Multi-Perspective Opinions Deliberation (P=2, 3, 4)

Parallel candidate reviews (Security, Performance, Maintainability, Reliability) generated simultaneously

All HTTP 200 OK

Perspectives	Assigned Evaluation Angles	Serialized Wall	Concurrent Wall	Speedup	Peak Slots	Candidate Statuses
P=2	Security, Performance	1.182 s	0.749 s	1.58×	2	[200, 200] (0 errors)
P=3	Security, Performance, Maintainability	1.460 s	1.259 s	1.16×	3	[200, 200, 200] (0 errors)
P=4	Security, Performance, Maintainability, Reliability	2.601 s	1.429 s	1.82×	4	[200, 200, 200, 200] (0 errors)

Architecture & Unified Memory Slot Distribution

Slot gate allocation and Metal unified memory residency

Hardware & Weights

Parameter sizing and physical unified memory requirements

Model Family Verdict Terse

Architecture Classification Sparse MoE

Total Parameters 30B

Active Parameters / Token ~3.0B

Resident Weight Memory 19.2 GB

Context & Slot Headroom

Multi-slot allocation configured under ENGINE_SLOT_MODE=model_aware

Total Loaded Context 262,144 tokens

Parallel Slots 4 active slots

Per-Slot Headroom 65,536 tokens / slot

Queue Latency Reduction 1.08s → 0.00s

Slot Gate Stability 100% (Zero Drop / Failures)

Qwen 3 Coder 30B (MLX)

MLX (MoE 30B-A3B) Qwen 3 Coder 4 Parallel Slots

Identifier: qwen/qwen3-coder-30b · Engine: MLX (Apple Silicon Metal)

Peak C=4 Speedup 1.37×

C=4 Concurrent Tok/s 72.5 tok/s

Throughput Gain +36.6%

Opinions P=4 Speedup 1.91×

Loaded Context 262,144 tokens

RAM Residency 19.5 GB

Direct Concurrency Scaling (C=1, 2, 3, 4)

Head-to-head comparison of Serialized execution (OFF) vs Model-Aware parallel slot execution (ON)

Aggregate Generation Throughput Scaling (tok/s) Serialized: 53.1 tok/s ➔ Concurrent: 72.5 tok/s (+36.6%)

Serialized (Concurrency OFF)

Concurrent (Concurrency ON, 4 Parallel Slots)

Load	Mode	Wall Time	Slots	Mean Latency	Mean Queue Wait	Throughput	Speedup	TP Gain
C=1	OFF (serialize)	0.652 s	1	0.652 s	0.000 s	75.1 tok/s	1.00×	1.00×
C=1	ON (model_aware)	0.581 s	1	0.581 s	0.01 ms	84.4 tok/s	1.12×	+12.3%
C=2	OFF (serialize)	2.312 s	1	1.156 s	0.335 s	42.4 tok/s	1.00×	1.00×
C=2	ON (model_aware)	1.586 s	2	1.280 s	0.01 ms	61.8 tok/s	1.46×	+45.8%
C=3	OFF (serialize)	2.539 s	1	0.846 s	0.747 s	57.9 tok/s	1.00×	1.00×
C=3	ON (model_aware)	1.767 s	3	1.308 s	0.01 ms	83.2 tok/s	1.44×	+43.7%
C=4	OFF (serialize)	3.693 s	1	0.923 s	1.344 s	53.1 tok/s	1.00×	1.00×
C=4	ON (model_aware)	2.705 s	4	2.166 s	0.01 ms	72.5 tok/s	1.37×	+36.6%

Multi-Perspective Opinions Deliberation (P=2, 3, 4)

Parallel candidate reviews (Security, Performance, Maintainability, Reliability) generated simultaneously

All HTTP 200 OK

Perspectives	Assigned Evaluation Angles	Serialized Wall	Concurrent Wall	Speedup	Peak Slots	Candidate Statuses
P=2	Security, Performance	1.123 s	0.744 s	1.51×	2	[200, 200] (0 errors)
P=3	Security, Performance, Maintainability	1.854 s	1.016 s	1.82×	3	[200, 200, 200] (0 errors)
P=4	Security, Performance, Maintainability, Reliability	2.636 s	1.383 s	1.91×	4	[200, 200, 200, 200] (0 errors)

Architecture & Unified Memory Slot Distribution

Slot gate allocation and Metal unified memory residency

Hardware & Weights

Parameter sizing and physical unified memory requirements

Model Family Qwen 3 Coder

Architecture Classification Sparse MoE

Total Parameters 30B

Active Parameters / Token ~3.0B

Resident Weight Memory 19.5 GB

Context & Slot Headroom

Multi-slot allocation configured under ENGINE_SLOT_MODE=model_aware

Total Loaded Context 262,144 tokens

Parallel Slots 4 active slots

Per-Slot Headroom 65,536 tokens / slot

Queue Latency Reduction 1.34s → 0.00s

Slot Gate Stability 100% (Zero Drop / Failures)

Verdict 4 4B Quill (MLX)

MLX (Dense 4B) Verdict Quill 4 Parallel Slots

Identifier: verdict-4-4b-quill-mlx · Engine: MLX (Apple Silicon Metal)

Peak C=4 Speedup 1.59×

C=4 Concurrent Tok/s 158.2 tok/s

Throughput Gain +58.8%

Opinions P=4 Speedup 1.87×

Loaded Context 262,144 tokens

RAM Residency 3.6 GB

Direct Concurrency Scaling (C=1, 2, 3, 4)

Head-to-head comparison of Serialized execution (OFF) vs Model-Aware parallel slot execution (ON)

Aggregate Generation Throughput Scaling (tok/s) Serialized: 99.6 tok/s ➔ Concurrent: 158.2 tok/s (+58.8%)

Serialized (Concurrency OFF)

Concurrent (Concurrency ON, 4 Parallel Slots)

Load	Mode	Wall Time	Slots	Mean Latency	Mean Queue Wait	Throughput	Speedup	TP Gain
C=1	OFF (serialize)	0.441 s	1	0.441 s	0.000 s	111.2 tok/s	1.00×	1.00×
C=1	ON (model_aware)	0.405 s	1	0.405 s	0.01 ms	120.9 tok/s	1.09×	+8.7%
C=2	OFF (serialize)	0.801 s	1	0.400 s	0.206 s	122.3 tok/s	1.00×	1.00×
C=2	ON (model_aware)	0.525 s	2	0.520 s	0.01 ms	186.7 tok/s	1.53×	+52.6%
C=3	OFF (serialize)	1.273 s	1	0.424 s	0.422 s	115.4 tok/s	1.00×	1.00×
C=3	ON (model_aware)	1.621 s	3	1.606 s	0.02 ms	90.7 tok/s	0.79×	+-21.4%
C=4	OFF (serialize)	1.967 s	1	0.492 s	0.696 s	99.6 tok/s	1.00×	1.00×
C=4	ON (model_aware)	1.239 s	4	1.212 s	0.01 ms	158.2 tok/s	1.59×	+58.8%

Multi-Perspective Opinions Deliberation (P=2, 3, 4)

Parallel candidate reviews (Security, Performance, Maintainability, Reliability) generated simultaneously

All HTTP 200 OK

Perspectives	Assigned Evaluation Angles	Serialized Wall	Concurrent Wall	Speedup	Peak Slots	Candidate Statuses
P=2	Security, Performance	0.674 s	0.507 s	1.33×	2	[200, 200] (0 errors)
P=3	Security, Performance, Maintainability	1.077 s	0.733 s	1.47×	3	[200, 200, 200] (0 errors)
P=4	Security, Performance, Maintainability, Reliability	2.235 s	1.194 s	1.87×	4	[200, 200, 200, 200] (0 errors)

Architecture & Unified Memory Slot Distribution

Slot gate allocation and Metal unified memory residency

Hardware & Weights

Parameter sizing and physical unified memory requirements

Model Family Verdict Quill

Architecture Classification Dense

Total Parameters 4B

Active Parameters / Token 4.0B

Resident Weight Memory 3.6 GB

Context & Slot Headroom

Multi-slot allocation configured under ENGINE_SLOT_MODE=model_aware

Total Loaded Context 262,144 tokens

Parallel Slots 4 active slots

Per-Slot Headroom 65,536 tokens / slot

Queue Latency Reduction 0.70s → 0.00s

Slot Gate Stability 100% (Zero Drop / Failures)