Datasets:
run_id stringlengths 24 24 | timestamp stringlengths 32 32 | system stringclasses 3
values | model_id stringclasses 1
value | model_revision stringclasses 1
value | quantization stringclasses 8
values | runtime stringclasses 5
values | runtime_revision stringclasses 4
values | mtp_enabled bool 2
classes | context_tokens int64 4.1k 262k | concurrency int64 1 1 | seed int64 0 33 | prompt_tokens int64 17 252k ⌀ | generated_tokens int64 3 1.02k ⌀ | ttft_seconds float64 0.12 1.66k ⌀ | prefill_tokens_per_second null | decode_tokens_per_second float64 16.1 4.49k ⌀ | aggregate_tokens_per_second null | peak_memory_gb null | average_power_watts null | energy_per_output_token null | mtp_proposed null | mtp_accepted null | mtp_acceptance_rate null | quality_task null | quality_score null | correctness_status stringclasses 2
values | error stringclasses 3
values | full_launch_command stringclasses 21
values | cache_state stringclasses 2
values | metric_scope stringclasses 1
value | per_request_ttft_seconds float64 0.12 1.66k ⌀ | per_request_decode_tokens_per_second float64 16.1 4.49k ⌀ | per_request_end_to_end_seconds float64 0 1.66k | output_corrupted bool 1
class | retry_count int64 0 0 | retry_reason null | kv_cache_enabled bool 1
class | average_accepted_draft_length null | schema_version stringclasses 1
value | sample_id stringlengths 31 31 | sequence int64 0 5 | status stringclasses 2
values | started_at stringlengths 32 32 | completed_at stringlengths 32 32 | config_sha256 stringlengths 64 64 | platform_id stringclasses 3
values | model stringclasses 1
value | task_category stringclasses 1
value | prompt_sha256 stringclasses 34
values | prompt_text stringclasses 34
values | raw_response dict | output_text stringclasses 69
values | finish_reason stringclasses 2
values | timing dict | usage dict | repetition int64 0 0 | error_type stringclasses 3
values | error_message stringclasses 3
values | environment_path stringlengths 53 53 | tags dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
run-0148a7ad6e0efe793a13 | 2026-07-12T14:28:24.571987+00:00 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | NVIDIA ModelOpt NVFP4 checkpoint, Marlin weight-only runtime fallback | vllm | vLLM main 1bd8f80a643efb5f977469692e458fd9a4fd3b4f + mapping patch 67d306c3eff39bffb654412b637b17e44f45db4a | true | 4,096 | 1 | 4 | 28 | 12 | 0.302063 | null | 92.308681 | null | null | null | null | null | null | null | null | null | not_evaluated | null | docker run --gpus all --ipc=host -p 8000:8000 qwen36-vllm-patched:main nvidia/Qwen3.6-35B-A3B-NVFP4 --revision 491c2f1ea524c639598bf8fa787a93fed5a6fbce --served-model-name qwen36-35b-a3b-nvfp4-mtp --trust-remote-code --max-model-len 4096 --gpu-memory-utilization 0.85 --max-num-seqs 1 --enforce-eager --speculative-confi... | cold_prompt | per_request | 0.302063 | 92.308681 | 0.432061 | false | 0 | null | true | null | 1.0 | sample-519dc4644c5f8b3d3177f162 | 0 | complete | 2026-07-12T14:28:24.139839+00:00 | 2026-07-12T14:28:24.571987+00:00 | 0148a7ad6e0efe793a1337caecc6d47a40eec967f2f499c54169656083566350 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | performance | bd2e9d3bb4700a1f1527a31b70f77ea0931d9b590bfaa9698aeaf45bf681e270 | What is the capital of France? End your response exactly with FINAL: Paris. | {
"chunks": [
{
"id": "chatcmpl-a3008ca2d7bb0e2c",
"object": "chat.completion.chunk",
"created": 1783866504,
"model": "qwen36-35b-a3b-nvfp4-mtp",
"choices": [
{
"index": 0,
"delta": {
"role": "assistant",
"content": ""
},
... | The capital of France is Paris. FINAL: Paris. | stop | {
"request_start_ns": 5864976786121006,
"first_token_ns": 5864977088183766,
"response_end_ns": 5864977218182374,
"ttft_ms": 302.06276,
"prompt_ms": null,
"decode_ms": 129.998608,
"end_to_end_ms": 432.061368
} | {
"prompt_tokens": 28,
"completion_tokens": 12,
"total_tokens": 40,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 92.3086807206428,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-0148a7ad6e0efe793a13/environment.json | {
"suite": "preserved-puzzle-quality-v1",
"device": "gx10",
"artifact_revision": "491c2f1ea524c639598bf8fa787a93fed5a6fbce",
"backend": "marlin-weight-only",
"mtp_method": "mtp",
"mtp_speculative_tokens": "3",
"stage": "final",
"seed": "4",
"context_label": "4k",
"quality_task": "puzzle-v1",
"thin... |
run-0148a7ad6e0efe793a13 | 2026-07-12T14:28:25.821459+00:00 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | NVIDIA ModelOpt NVFP4 checkpoint, Marlin weight-only runtime fallback | vllm | vLLM main 1bd8f80a643efb5f977469692e458fd9a4fd3b4f + mapping patch 67d306c3eff39bffb654412b637b17e44f45db4a | true | 4,096 | 1 | 4 | 36 | 91 | 0.133473 | null | 81.73966 | null | null | null | null | null | null | null | null | null | not_evaluated | null | docker run --gpus all --ipc=host -p 8000:8000 qwen36-vllm-patched:main nvidia/Qwen3.6-35B-A3B-NVFP4 --revision 491c2f1ea524c639598bf8fa787a93fed5a6fbce --served-model-name qwen36-35b-a3b-nvfp4-mtp --trust-remote-code --max-model-len 4096 --gpu-memory-utilization 0.85 --max-num-seqs 1 --enforce-eager --speculative-confi... | cold_prompt | per_request | 0.133473 | 81.73966 | 1.246764 | false | 0 | null | true | null | 1.0 | sample-a53574c71840298e0f9ad0e6 | 1 | complete | 2026-07-12T14:28:24.574618+00:00 | 2026-07-12T14:28:25.821459+00:00 | 0148a7ad6e0efe793a1337caecc6d47a40eec967f2f499c54169656083566350 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | performance | b39fcba2eb8a37984a87be9444c0248458b30bdf9fd076507984f1659dd25ef6 | Write a Python function fibonacci(n) returning the first n Fibonacci numbers starting with 0. Return only executable Python code. | {
"chunks": [
{
"id": "chatcmpl-b03e7b3da419b7f6",
"object": "chat.completion.chunk",
"created": 1783866504,
"model": "qwen36-35b-a3b-nvfp4-mtp",
"choices": [
{
"index": 0,
"delta": {
"role": "assistant",
"content": ""
},
... | ```python
def fibonacci(n):
if n <= 0:
return []
if n == 1:
return [0]
fib_sequence = [0, 1]
for i in range(2, n):
next_fib = fib_sequence[-1] + fib_sequence[-2]
fib_sequence.append(next_fib)
return fib_sequence
``` | stop | {
"request_start_ns": 5864977220900085,
"first_token_ns": 5864977354373495,
"response_end_ns": 5864978467664148,
"ttft_ms": 133.47341,
"prompt_ms": null,
"decode_ms": 1113.290653,
"end_to_end_ms": 1246.764063
} | {
"prompt_tokens": 36,
"completion_tokens": 91,
"total_tokens": 127,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 81.73966048738578,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-0148a7ad6e0efe793a13/environment.json | {
"suite": "preserved-puzzle-quality-v1",
"device": "gx10",
"artifact_revision": "491c2f1ea524c639598bf8fa787a93fed5a6fbce",
"backend": "marlin-weight-only",
"mtp_method": "mtp",
"mtp_speculative_tokens": "3",
"stage": "final",
"seed": "4",
"context_label": "4k",
"quality_task": "puzzle-v1",
"thin... |
run-0148a7ad6e0efe793a13 | 2026-07-12T14:28:29.384973+00:00 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | NVIDIA ModelOpt NVFP4 checkpoint, Marlin weight-only runtime fallback | vllm | vLLM main 1bd8f80a643efb5f977469692e458fd9a4fd3b4f + mapping patch 67d306c3eff39bffb654412b637b17e44f45db4a | true | 4,096 | 1 | 4 | 56 | 256 | 0.142014 | null | 74.873419 | null | null | null | null | null | null | null | null | null | not_evaluated | null | docker run --gpus all --ipc=host -p 8000:8000 qwen36-vllm-patched:main nvidia/Qwen3.6-35B-A3B-NVFP4 --revision 491c2f1ea524c639598bf8fa787a93fed5a6fbce --served-model-name qwen36-35b-a3b-nvfp4-mtp --trust-remote-code --max-model-len 4096 --gpu-memory-utilization 0.85 --max-num-seqs 1 --enforce-eager --speculative-confi... | cold_prompt | per_request | 0.142014 | 74.873419 | 3.561118 | false | 0 | null | true | null | 1.0 | sample-4b5d0f731ff309a8dbe41d62 | 2 | complete | 2026-07-12T14:28:25.823787+00:00 | 2026-07-12T14:28:29.384973+00:00 | 0148a7ad6e0efe793a1337caecc6d47a40eec967f2f499c54169656083566350 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | performance | b83985c659380d1d4080e3a271ae9e38465e06dfac3401644205d58e3f24463a | Three machines make three parts in three minutes at identical constant rates. How many minutes do 100 machines need to make 2,000 parts? End your response exactly with FINAL: 60. | {
"chunks": [
{
"id": "chatcmpl-b8bffb65a62e84f6",
"object": "chat.completion.chunk",
"created": 1783866505,
"model": "qwen36-35b-a3b-nvfp4-mtp",
"choices": [
{
"index": 0,
"delta": {
"role": "assistant",
"content": ""
},
... | To solve this problem, we first need to determine the production rate of a single machine.
**Step 1: Determine the rate of one machine.**
The problem states that 3 machines make 3 parts in 3 minutes.
Since the machines work at identical constant rates, we can isolate the work done by one machine.
If 3 machines make ... | length | {
"request_start_ns": 5864978470067064,
"first_token_ns": 5864978612081363,
"response_end_ns": 5864982031185273,
"ttft_ms": 142.014299,
"prompt_ms": null,
"decode_ms": 3419.10391,
"end_to_end_ms": 3561.118209
} | {
"prompt_tokens": 56,
"completion_tokens": 256,
"total_tokens": 312,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 74.87341910003548,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-0148a7ad6e0efe793a13/environment.json | {
"suite": "preserved-puzzle-quality-v1",
"device": "gx10",
"artifact_revision": "491c2f1ea524c639598bf8fa787a93fed5a6fbce",
"backend": "marlin-weight-only",
"mtp_method": "mtp",
"mtp_speculative_tokens": "3",
"stage": "final",
"seed": "4",
"context_label": "4k",
"quality_task": "puzzle-v1",
"thin... |
run-0148a7ad6e0efe793a13 | 2026-07-12T14:28:29.837922+00:00 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | NVIDIA ModelOpt NVFP4 checkpoint, Marlin weight-only runtime fallback | vllm | vLLM main 1bd8f80a643efb5f977469692e458fd9a4fd3b4f + mapping patch 67d306c3eff39bffb654412b637b17e44f45db4a | true | 4,096 | 1 | 4 | 49 | 28 | 0.135621 | null | 89.007288 | null | null | null | null | null | null | null | null | null | not_evaluated | null | docker run --gpus all --ipc=host -p 8000:8000 qwen36-vllm-patched:main nvidia/Qwen3.6-35B-A3B-NVFP4 --revision 491c2f1ea524c639598bf8fa787a93fed5a6fbce --served-model-name qwen36-35b-a3b-nvfp4-mtp --trust-remote-code --max-model-len 4096 --gpu-memory-utilization 0.85 --max-num-seqs 1 --enforce-eager --speculative-confi... | cold_prompt | per_request | 0.135621 | 89.007288 | 0.450202 | false | 0 | null | true | null | 1.0 | sample-b586f615aa1936227ff3f5e1 | 3 | complete | 2026-07-12T14:28:29.387655+00:00 | 2026-07-12T14:28:29.837922+00:00 | 0148a7ad6e0efe793a1337caecc6d47a40eec967f2f499c54169656083566350 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | performance | 2b65b72640deff2f98daa0773f75a7253c29634b4002b2eef65b02e01bd0d6c1 | Return only valid JSON with exactly these values and no additional keys: project is "atlas", priority is 3, ready is true, and owners are ["Ada", "Lin"]. | {
"chunks": [
{
"id": "chatcmpl-8d3f8c980d63031c",
"object": "chat.completion.chunk",
"created": 1783866509,
"model": "qwen36-35b-a3b-nvfp4-mtp",
"choices": [
{
"index": 0,
"delta": {
"role": "assistant",
"content": ""
},
... | {"project": "atlas", "priority": 3, "ready": true, "owners": ["Ada", "Lin"]} | stop | {
"request_start_ns": 5864982033935886,
"first_token_ns": 5864982169557360,
"response_end_ns": 5864982484138340,
"ttft_ms": 135.621474,
"prompt_ms": null,
"decode_ms": 314.58098,
"end_to_end_ms": 450.202454
} | {
"prompt_tokens": 49,
"completion_tokens": 28,
"total_tokens": 77,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 89.0072883618075,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-0148a7ad6e0efe793a13/environment.json | {
"suite": "preserved-puzzle-quality-v1",
"device": "gx10",
"artifact_revision": "491c2f1ea524c639598bf8fa787a93fed5a6fbce",
"backend": "marlin-weight-only",
"mtp_method": "mtp",
"mtp_speculative_tokens": "3",
"stage": "final",
"seed": "4",
"context_label": "4k",
"quality_task": "puzzle-v1",
"thin... |
run-0148a7ad6e0efe793a13 | 2026-07-12T14:28:30.198432+00:00 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | NVIDIA ModelOpt NVFP4 checkpoint, Marlin weight-only runtime fallback | vllm | vLLM main 1bd8f80a643efb5f977469692e458fd9a4fd3b4f + mapping patch 67d306c3eff39bffb654412b637b17e44f45db4a | true | 4,096 | 1 | 4 | 48 | 18 | 0.135422 | null | 80.842038 | null | null | null | null | null | null | null | null | null | not_evaluated | null | docker run --gpus all --ipc=host -p 8000:8000 qwen36-vllm-patched:main nvidia/Qwen3.6-35B-A3B-NVFP4 --revision 491c2f1ea524c639598bf8fa787a93fed5a6fbce --served-model-name qwen36-35b-a3b-nvfp4-mtp --trust-remote-code --max-model-len 4096 --gpu-memory-utilization 0.85 --max-num-seqs 1 --enforce-eager --speculative-confi... | cold_prompt | per_request | 0.135422 | 80.842038 | 0.358079 | false | 0 | null | true | null | 1.0 | sample-7e4907e55dfe88e916cc2346 | 4 | complete | 2026-07-12T14:28:29.840287+00:00 | 2026-07-12T14:28:30.198432+00:00 | 0148a7ad6e0efe793a1337caecc6d47a40eec967f2f499c54169656083566350 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | performance | f06d19b46df492425f843d0b14128946627defb7c06eef709446577917aa5fac | Represent this tool call as JSON only: call weather for Tokyo using celsius. Use exactly {"tool":"weather","arguments":{"city":"Tokyo","unit":"celsius"}}. | {
"chunks": [
{
"id": "chatcmpl-8834d35e1b5620cf",
"object": "chat.completion.chunk",
"created": 1783866509,
"model": "qwen36-35b-a3b-nvfp4-mtp",
"choices": [
{
"index": 0,
"delta": {
"role": "assistant",
"content": ""
},
... | {"tool":"weather","arguments":{"city":"Tokyo","unit":"celsius"}} | stop | {
"request_start_ns": 5864982486567816,
"first_token_ns": 5864982621989994,
"response_end_ns": 5864982844646429,
"ttft_ms": 135.422178,
"prompt_ms": null,
"decode_ms": 222.656435,
"end_to_end_ms": 358.078613
} | {
"prompt_tokens": 48,
"completion_tokens": 18,
"total_tokens": 66,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 80.84203809335222,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-0148a7ad6e0efe793a13/environment.json | {
"suite": "preserved-puzzle-quality-v1",
"device": "gx10",
"artifact_revision": "491c2f1ea524c639598bf8fa787a93fed5a6fbce",
"backend": "marlin-weight-only",
"mtp_method": "mtp",
"mtp_speculative_tokens": "3",
"stage": "final",
"seed": "4",
"context_label": "4k",
"quality_task": "puzzle-v1",
"thin... |
run-0148a7ad6e0efe793a13 | 2026-07-12T14:28:30.879914+00:00 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | NVIDIA ModelOpt NVFP4 checkpoint, Marlin weight-only runtime fallback | vllm | vLLM main 1bd8f80a643efb5f977469692e458fd9a4fd3b4f + mapping patch 67d306c3eff39bffb654412b637b17e44f45db4a | true | 4,096 | 1 | 4 | 1,244 | 30 | 0.340935 | null | 88.886868 | null | null | null | null | null | null | null | null | null | not_evaluated | null | docker run --gpus all --ipc=host -p 8000:8000 qwen36-vllm-patched:main nvidia/Qwen3.6-35B-A3B-NVFP4 --revision 491c2f1ea524c639598bf8fa787a93fed5a6fbce --served-model-name qwen36-35b-a3b-nvfp4-mtp --trust-remote-code --max-model-len 4096 --gpu-memory-utilization 0.85 --max-num-seqs 1 --enforce-eager --speculative-confi... | cold_prompt | per_request | 0.340935 | 88.886868 | 0.678443 | false | 0 | null | true | null | 1.0 | sample-ec30b5340a037479fec2d8d1 | 5 | complete | 2026-07-12T14:28:30.201404+00:00 | 2026-07-12T14:28:30.879914+00:00 | 0148a7ad6e0efe793a1337caecc6d47a40eec967f2f499c54169656083566350 | asus-gx10 | Qwen/Qwen3.6-35B-A3B | performance | 579e79746cb56f008e79732fd790cd7ed027eabc1a95904033632c86e247dc31 | Read the records below and identify the deployment authorization code in Record 073. End your response exactly with FINAL: <code>.
Record 000: value-000
Record 001: value-001
Record 002: value-002
Record 003: value-003
Record 004: value-004
Record 005: value-005
Record 006: value-006
Record 007: value-007
Record 008: ... | {
"chunks": [
{
"id": "chatcmpl-b96d9576612c5938",
"object": "chat.completion.chunk",
"created": 1783866510,
"model": "qwen36-35b-a3b-nvfp4-mtp",
"choices": [
{
"index": 0,
"delta": {
"role": "assistant",
"content": ""
},
... | The deployment authorization code in Record 073 is KITE-7391.
FINAL: KITE-7391 | stop | {
"request_start_ns": 5864982847684476,
"first_token_ns": 5864983188619732,
"response_end_ns": 5864983526127404,
"ttft_ms": 340.935256,
"prompt_ms": null,
"decode_ms": 337.507672,
"end_to_end_ms": 678.442928
} | {
"prompt_tokens": 1244,
"completion_tokens": 30,
"total_tokens": 1274,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 88.88686832576653,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-0148a7ad6e0efe793a13/environment.json | {
"suite": "preserved-puzzle-quality-v1",
"device": "gx10",
"artifact_revision": "491c2f1ea524c639598bf8fa787a93fed5a6fbce",
"backend": "marlin-weight-only",
"mtp_method": "mtp",
"mtp_speculative_tokens": "3",
"stage": "final",
"seed": "4",
"context_label": "4k",
"quality_task": "puzzle-v1",
"thin... |
run-01bbb7e9ed5688591fba | 2026-07-11T20:22:16.595420+00:00 | apple-m2-max-64gb | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | MLX affine 6-bit, group size 64, selected gates 8-bit | mlx-lm | a9d4e2b57679ea2d892c646c019230152af30901 | false | 4,096 | 1 | 33 | 3,098 | 28 | 0.244929 | null | 58.323978 | null | null | null | null | null | null | null | null | null | not_evaluated | null | .venv/bin/python -m mlx_lm server --model ~/Models/Qwen3.6-35B-A3B-MLX-6bit --host 0.0.0.0 --port 8080 --temp 0 --top-p 1 --max-tokens 256 --chat-template-args '{"enable_thinking":false}' --decode-concurrency 4 --prompt-concurrency 4 --log-level INFO | cold_prompt | per_request | 0.244929 | 58.323978 | 0.725006 | false | 0 | null | true | null | 1.0 | sample-f325dc6a095e07839abbafa0 | 0 | complete | 2026-07-11T20:22:15.870323+00:00 | 2026-07-11T20:22:16.595420+00:00 | 01bbb7e9ed5688591fba6e730bbe808cb6090a4a8d3de1a1568c2b831290337b | apple-m2-max-64gb | Qwen/Qwen3.6-35B-A3B | performance | 6bd9c70e66f45e8076ed2f147b316b588f37f802e18040fa857d03303ddd605e | DOCUMENTS
[d00] Archive record 00. This record discusses routine inventory, maintenance windows, and quarterly planning. Reference values 1000 and zone-0 are unrelated to user questions. Archive record 00. This record discusses routine inventory, maintenance windows, and quarterly planning. Reference values 1000 and zo... | {
"chunks": [
{
"id": "chatcmpl-26736d15-c51b-4554-bf42-4a05f5784f3c",
"object": "chat.completion.chunk",
"created": 1783801335,
"model": "/Users/benchmark/Models/Qwen3.6-35B-A3B-MLX-6bit",
"choices": [
{
"index": 0,
"delta": {
"role": "assista... | {
"answer": "LIME-482",
"citations": [
"d17"
]
} | stop | {
"request_start_ns": 5799808516608756,
"first_token_ns": 5799808761537963,
"response_end_ns": 5799809241614955,
"ttft_ms": 244.929207,
"prompt_ms": null,
"decode_ms": 480.076992,
"end_to_end_ms": 725.006199
} | {
"prompt_tokens": 3098,
"completion_tokens": 28,
"total_tokens": 3126,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 58.3239781672353,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-01bbb7e9ed5688591fba/environment.json | {
"suite": "stage2-retrieval",
"device": "m2-max",
"artifact_revision": "cb7e092ef8efe540bc3672c8929c4adbe5f4f759",
"backend": null,
"mtp_method": null,
"mtp_speculative_tokens": null,
"stage": "2",
"seed": "33",
"context_label": "4k",
"quality_task": "retrieval-v1",
"thinking": "disabled",
"qua... |
run-01bbb7e9ed5688591fba | 2026-07-11T20:22:25.470591+00:00 | apple-m2-max-64gb | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | MLX affine 6-bit, group size 64, selected gates 8-bit | mlx-lm | a9d4e2b57679ea2d892c646c019230152af30901 | false | 4,096 | 1 | 33 | 3,032 | 26 | 8.354609 | null | 50.266023 | null | null | null | null | null | null | null | null | null | not_evaluated | null | .venv/bin/python -m mlx_lm server --model ~/Models/Qwen3.6-35B-A3B-MLX-6bit --host 0.0.0.0 --port 8080 --temp 0 --top-p 1 --max-tokens 256 --chat-template-args '{"enable_thinking":false}' --decode-concurrency 4 --prompt-concurrency 4 --log-level INFO | cold_prompt | per_request | 8.354609 | 50.266023 | 8.871857 | false | 0 | null | true | null | 1.0 | sample-e9d8e4eca9c48da16cdf4297 | 1 | complete | 2026-07-11T20:22:16.598659+00:00 | 2026-07-11T20:22:25.470591+00:00 | 01bbb7e9ed5688591fba6e730bbe808cb6090a4a8d3de1a1568c2b831290337b | apple-m2-max-64gb | Qwen/Qwen3.6-35B-A3B | performance | 7aca262bf03f479f14a4d157863b51ff13b9161b007045044e442b86a3ec6e53 | DOCUMENTS
[d00] Archive record 00. This record discusses routine inventory, maintenance windows, and quarterly planning. Reference values 1000 and zone-0 are unrelated to user questions. Archive record 00. This record discusses routine inventory, maintenance windows, and quarterly planning. Reference values 1000 and zo... | {
"chunks": [
{
"id": "chatcmpl-e3d78063-d374-47e9-b9e0-3337846388bd",
"object": "chat.completion.chunk",
"created": 1783801336,
"model": "/Users/benchmark/Models/Qwen3.6-35B-A3B-MLX-6bit",
"choices": [
{
"index": 0,
"delta": {
"role": "assista... | {
"answer": "7319",
"citations": [
"d23"
]
} | stop | {
"request_start_ns": 5799809244939634,
"first_token_ns": 5799817599549089,
"response_end_ns": 5799818116797096,
"ttft_ms": 8354.609455,
"prompt_ms": null,
"decode_ms": 517.248007,
"end_to_end_ms": 8871.857462
} | {
"prompt_tokens": 3032,
"completion_tokens": 26,
"total_tokens": 3058,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 50.26602258131078,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-01bbb7e9ed5688591fba/environment.json | {
"suite": "stage2-retrieval",
"device": "m2-max",
"artifact_revision": "cb7e092ef8efe540bc3672c8929c4adbe5f4f759",
"backend": null,
"mtp_method": null,
"mtp_speculative_tokens": null,
"stage": "2",
"seed": "33",
"context_label": "4k",
"quality_task": "retrieval-v1",
"thinking": "disabled",
"qua... |
run-01bbb7e9ed5688591fba | 2026-07-11T20:22:34.885195+00:00 | apple-m2-max-64gb | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | MLX affine 6-bit, group size 64, selected gates 8-bit | mlx-lm | a9d4e2b57679ea2d892c646c019230152af30901 | false | 4,096 | 1 | 33 | 3,032 | 29 | 8.83699 | null | 50.447742 | null | null | null | null | null | null | null | null | null | not_evaluated | null | .venv/bin/python -m mlx_lm server --model ~/Models/Qwen3.6-35B-A3B-MLX-6bit --host 0.0.0.0 --port 8080 --temp 0 --top-p 1 --max-tokens 256 --chat-template-args '{"enable_thinking":false}' --decode-concurrency 4 --prompt-concurrency 4 --log-level INFO | cold_prompt | per_request | 8.83699 | 50.447742 | 9.411842 | false | 0 | null | true | null | 1.0 | sample-50c98ab6450ab72006ac26f2 | 2 | complete | 2026-07-11T20:22:25.473271+00:00 | 2026-07-11T20:22:34.885195+00:00 | 01bbb7e9ed5688591fba6e730bbe808cb6090a4a8d3de1a1568c2b831290337b | apple-m2-max-64gb | Qwen/Qwen3.6-35B-A3B | performance | 28af6065de9bb49b756195acd824b703d9a9f3036a5475a97d96fb01aadae456 | DOCUMENTS
[d00] Archive record 00. This record discusses routine inventory, maintenance windows, and quarterly planning. Reference values 1000 and zone-0 are unrelated to user questions. Archive record 00. This record discusses routine inventory, maintenance windows, and quarterly planning. Reference values 1000 and zo... | {
"chunks": [
{
"id": "chatcmpl-7294f06a-b5d3-458e-b470-5d2820248d66",
"object": "chat.completion.chunk",
"created": 1783801345,
"model": "/Users/benchmark/Models/Qwen3.6-35B-A3B-MLX-6bit",
"choices": [
{
"index": 0,
"delta": {
"role": "assista... | {
"answer": "violet",
"citations": [
"d08",
"d31"
]
} | stop | {
"request_start_ns": 5799818119551236,
"first_token_ns": 5799826956540748,
"response_end_ns": 5799827531393033,
"ttft_ms": 8836.989512,
"prompt_ms": null,
"decode_ms": 574.852285,
"end_to_end_ms": 9411.841797
} | {
"prompt_tokens": 3032,
"completion_tokens": 29,
"total_tokens": 3061,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 50.44774241438389,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-01bbb7e9ed5688591fba/environment.json | {
"suite": "stage2-retrieval",
"device": "m2-max",
"artifact_revision": "cb7e092ef8efe540bc3672c8929c4adbe5f4f759",
"backend": null,
"mtp_method": null,
"mtp_speculative_tokens": null,
"stage": "2",
"seed": "33",
"context_label": "4k",
"quality_task": "retrieval-v1",
"thinking": "disabled",
"qua... |
run-01bbb7e9ed5688591fba | 2026-07-11T20:22:44.720914+00:00 | apple-m2-max-64gb | Qwen/Qwen3.6-35B-A3B | 995ad96eacd98c81ed38be0c5b274b04031597b0 | MLX affine 6-bit, group size 64, selected gates 8-bit | mlx-lm | a9d4e2b57679ea2d892c646c019230152af30901 | false | 4,096 | 1 | 33 | 3,146 | 16 | 9.508664 | null | 49.447405 | null | null | null | null | null | null | null | null | null | not_evaluated | null | .venv/bin/python -m mlx_lm server --model ~/Models/Qwen3.6-35B-A3B-MLX-6bit --host 0.0.0.0 --port 8080 --temp 0 --top-p 1 --max-tokens 256 --chat-template-args '{"enable_thinking":false}' --decode-concurrency 4 --prompt-concurrency 4 --log-level INFO | cold_prompt | per_request | 9.508664 | 49.447405 | 9.83224 | false | 0 | null | true | null | 1.0 | sample-518b0dcc0be4b5acadc78fa1 | 3 | complete | 2026-07-11T20:22:34.888605+00:00 | 2026-07-11T20:22:44.720914+00:00 | 01bbb7e9ed5688591fba6e730bbe808cb6090a4a8d3de1a1568c2b831290337b | apple-m2-max-64gb | Qwen/Qwen3.6-35B-A3B | performance | f53d2e953e47f29372ea2f383d45a39998528e790bd7944c20f6400cc5f8a7f3 | DOCUMENTS
[d00] Archive record 00. This record discusses routine inventory, maintenance windows, and quarterly planning. Reference values 1000 and zone-0 are unrelated to user questions. Archive record 00. This record discusses routine inventory, maintenance windows, and quarterly planning. Reference values 1000 and zo... | {
"chunks": [
{
"id": "chatcmpl-4acba0f2-ffe4-4828-ad9b-a0675b7c4694",
"object": "chat.completion.chunk",
"created": 1783801354,
"model": "/Users/benchmark/Models/Qwen3.6-35B-A3B-MLX-6bit",
"choices": [
{
"index": 0,
"delta": {
"role": "assista... | {
"answer": "UNKNOWN",
"citations": []
} | stop | {
"request_start_ns": 5799827534886983,
"first_token_ns": 5799837043550613,
"response_end_ns": 5799837367126745,
"ttft_ms": 9508.66363,
"prompt_ms": null,
"decode_ms": 323.576132,
"end_to_end_ms": 9832.239762
} | {
"prompt_tokens": 3146,
"completion_tokens": 16,
"total_tokens": 3162,
"prompt_tokens_per_second": null,
"decode_tokens_per_second": 49.44740485370534,
"aggregate_tokens_per_second": null
} | 0 | null | null | raw-results/run-01bbb7e9ed5688591fba/environment.json | {
"suite": "stage2-retrieval",
"device": "m2-max",
"artifact_revision": "cb7e092ef8efe540bc3672c8929c4adbe5f4f759",
"backend": null,
"mtp_method": null,
"mtp_speculative_tokens": null,
"stage": "2",
"seed": "33",
"context_label": "4k",
"quality_task": "retrieval-v1",
"thinking": "disabled",
"qua... |
End of preview. Expand in Data Studio
Qwen3.6-35B-A3B cross-platform benchmark dataset
Complete sanitized evidence for the Qwen3.6 GX10 versus M2 Max benchmark.
Headline results
- At 128K, median cold-prompt TTFT was 45.70 s on GX10 and 552.79 s on M2 Max, a 12.10x difference.
- Near 256K, GX10 completed and strictly passed 12/12 requests. M2 Max completed 9/12 and strictly passed 7/9 completed requests.
- On the same Q4_K_M coding control, MTP improved median decode 26.87% on GX10 CUDA and regressed 19.35% on M2 Max Metal.
Read the final technical report for configuration details, quality denominators, runtime failures, and limitations.
Contents
processed-results/results.parquet: 429 normalized result rows for analysis.processed-results/results.jsonl: the same generated table as JSONL.raw-results/: append-only outputs, resolved configs, and environment records.configs/: staged benchmark configurations.tasks/: deterministic coding, tool, retrieval, and Puzzle contracts.reports/: final reports and the agent working log.reproductions/: retained vLLM and llama.cpp failure evidence.figures/: charts generated from the raw dataset.
Privacy transformation
The exact private archive is retained offline. This public dataset replaces local macOS/Linux usernames, personal home paths, hostnames, and Tailscale IP addresses with stable pseudonymous placeholders. Model outputs, timings, task inputs, quality results, model revisions, runtime revisions, and launch flags are unchanged.
Run IDs and sample IDs are preserved so published reports and row references remain stable. Because endpoint and filesystem fields were redacted after collection, a public config.resolved.json will not reproduce the original config hash byte-for-byte. Those fields are transport metadata, not benchmark variables.
Reproducibility
git clone https://github.com/sxuff/qwen36-cross-platform
cd qwen36-cross-platform
python -m pip install -e '.[dataset]'
python scripts/fetch_dataset.py --output data/
No model weights are included. Use the immutable upstream model and checkpoint revisions recorded in the reports.
License
Original benchmark code and materials are Apache-2.0. Model checkpoints and third-party runtimes retain their upstream licenses.
- Downloads last month
- 668