General Q&A · Coding

DeepSeek-V4.1-Flash-1017

Four MegaCube GB10 devices form one TP4 replica. After tuning, mean aggregate throughput across eight task categories is 171.66 token/s at C8; C1 median time to first token is 0.304 seconds.

Multimodal understandingText, images → text

4 MegaCubesPost-tuning testTested on 2026-10-19

01 / Model & scope

Model and validation scope

This record covers the final DSpark5 plan running under the cluster management platform: seven concurrency levels, cold-prompt prefill, memory sampling, basic functions and long-context retrieval. This is one v41bench v1 run, not the archived three-round real40 protocol; no direct speedup or previous quality score is carried forward.

Local Q&A and code generation
Tool calls and structured output
Fact retrieval from long inputs

Model version: DeepSeek-V4.1-Flash-1017
Weights and storage: Official native 1017 weights; full copies on each node’s local NVMe; no additional EXL3 conversion
Input / output: Text and basic image input; text, structured results and tool-call output
Context coverage: Configured limit: 300,000 tokens; single-request middle-position retrieval tested at 130,258 / 260,119 tokens
Data source: Post-tuning report dated 2026-10-19; compiled at 19:40:40 (UTC+08:00)

After a normal platform stop/start, all 17 basic functional checks passed. The original vision/tool suite passed 7/7 cases, retaining its images, prompts and rules while adapting only authentication, port and model alias. These limited checks are not a general answer or industry-vision accuracy score.

View tuned benchmark data (JSON) ↓

Evidence checked on 2026-10-19

02 / Deployment

MegaCube deployment configuration

Compute and topology: 4 × MegaCube GB10; one TP4 replica
Storage and interconnect: Complete weights on each node’s local NVMe; RoCE API1; streaming weight loading retained
Inference and prefill: DSpark=5; prefill batch tokens=8192; GMU=0.80
Scheduling and memory limit: Context 300,000 tokens; max sequences 8; Docker limit 112 GiB per node
Engram configuration: FAST=1; threads=128; chunk=16; retention=128
Communication and indexer: NCCL_MAX_NCHANNELS=8; AR≤2MB / AG≤16MB; indexer TP-split=1
Memory policy: Memory monitoring only; fixed 12 GiB container auto-stop rule removed

Verify weights and plan
Pin the official 1017 weights, reference v3 image and DSpark5 plan across all four nodes.
Prepare nodes and interconnect
Verify complete local NVMe weight copies, RoCE connectivity and the current plan parameters.
Check platform release coverage
This plan passed managed operation and normal stop/start checks; consult the release manifest to confirm inclusion in a public installer.
Re-test actual workloads
Re-test target tasks, APIs and streaming, including concurrency, context length and memory headroom.

This page records the tuned validation; neither legacy scripts nor the platform 1.1.0 installer proves that this plan is included.

Pinned weights and runtime versions

Pinned weight revision: 1bae0e1088993b9b82de08f59de5be5a34e42170
Container image: megacube/deepseek411017:reference-9fa230d-b58f34e-v3
Plan revision: deepseek411017-reference-roce-dspark5-p8192-gmu80-v3
CUDA graph: FULL_AND_PIECEWISE; 5, 6, 10, 12, 15, 18, 20, 24, 25, 30, 35, 36, 40, 42, 48
Allocator setting: expandable_segments:False
Benchmark script SHA256: e0d6b2d25bd585d11fbdf39c2ddcdf7a4de8ab685af6bd42465e69f3ee6e80a8

03 / Performance

MegaCube performance test results

2026-10-19 · 4 MegaCube nodes · v41bench v1 / managed-v1 · eight fixed task prompts · C1 / C2 / C3 / C4 / C5 / C6 / C8 · one run, non-thinking, SSE streaming.

171.66 token/sC8 · Mean aggregate throughput across eight categories

0.304 sC1 · Formal-request TTFT P50

268 / 268Complete SSE requests · not accuracy

Throughput, per-stream decode and mean TTFT are separate arithmetic averages of the eight formal task categories. P50/P90 pool all formal requests at each concurrency; counting tasks are excluded. Batch throughput = total output tokens / batch duration; stream decode = (output tokens − 1) / (duration − TTFT). Aggregate throughput divided by concurrency is not the per-stream decode metric.

Eight-category means and pooled-request quantiles are separate statistics, not three-round medians.

Seven concurrency levels
Concurrency	Mean aggregate token/s	Mean stream decode token/s	Mean TTFT s	TTFT P50 s	TTFT P90 s	Formal requests
C1	47.74	53.99	0.318	0.304	0.382	8
C2	76.71	44.07	0.375	0.332	0.395	16
C3	96.89	37.40	0.401	0.353	0.648	24
C4	117.11	34.29	0.414	0.374	0.731	32
C5	134.73	32.15	0.443	0.392	0.879	40
C6	152.88	29.86	0.492	0.409	0.712	48
C8	171.66	25.57	0.605	0.466	1.300	64

Uses actual input tokens. Prefill rate = input tokens / TTFT, not output throughput.

Cold-prompt prefill
Actual input tokens	TTFT s	Prefill token/s
2950	1.824	1617.1
11592	6.950	1668.0
46810	24.296	1926.6
93335	55.058	1695.2

One trial at each length, both at depth 0.5; not a multi-position, multi-round or concurrent long-context test.

Long-context retrieval
Actual input tokens	Needle depth	TTFT s	Prefill token/s	Answer	Result
130258	0.5	88.9	1465.6	COPPER-LANTERN-8315	Passed
260119	0.5	244.4	1064.5	COPPER-LANTERN-8315	Passed

Resource sampling details

Nodes are anonymized as 1–4; sampling is approximately once per second. Host used = Total − Available; the 112 GiB Docker limit is not actual usage.

Resource sampling details
Window	Node	Samples	Min available GiB	Max host used GiB	Max cgroup GiB
Formal performance	1	431	13.440	108.188	20.199
Formal performance	2	431	12.636	107.027	18.583
Formal performance	3	430	14.312	107.314	20.366
Formal performance	4	430	14.055	107.602	20.066
First managed start	1	346	15.064	106.563	35.710
First managed start	2	346	12.711	106.951	34.259
First managed start	3	347	13.719	107.907	35.244
First managed start	4	347	14.133	107.524	35.264
Platform restart	1	337	14.595	107.032	34.820
Platform restart	2	337	14.106	105.556	33.961
Platform restart	3	338	15.207	106.418	33.546
Platform restart	4	337	15.084	106.573	35.304
Pre-restart functions	1	12	13.971	107.656	20.125
Pre-restart functions	2	11	13.313	106.350	18.493
Pre-restart functions	3	11	14.726	106.900	19.880
Pre-restart functions	4	12	14.806	106.851	19.353
Post-restart functions	1	12	15.762	105.865	20.176
Post-restart functions	2	12	15.403	104.259	16.034
Post-restart functions	3	11	16.556	105.070	19.811
Post-restart functions	4	12	16.374	105.283	18.381
Final long-context retrieval	1	348	13.156	108.471	20.410
Final long-context retrieval	2	348	12.174	107.488	18.107
Final long-context retrieval	3	347	13.903	107.722	20.607
Final long-context retrieval	4	347	14.028	107.629	19.762

17 basic functional checks

Passed after a normal platform stop/start. Thinking output is a functional check, not the generation mode used for performance testing.

17 basic functional checks
Check	Result
api-authentication	Passed
chinese	Passed
math	Passed
english	Passed
json	Passed
code	Passed
tool-call	Passed
tool-result	Passed
vision-red	Passed
vision-four-images	Passed
vision-tool-call	Passed
sse-chinese	Passed
prefix-0	Passed
prefix-1	Passed
prefix-2	Passed
prefix-cache-hit-counters	Passed
thinking	Passed

7 vision and tool cases

7 vision and tool cases
Case	Result
V1 stripes	Passed
V2 two images	Passed
V3 grid	Passed
T1 tool call	Passed
T2 tool roundtrip	Passed
T3 parallel calls	Passed
T4 forced tool_choice	Passed

Request integrity audit

Covers all 268 formal, counting, warmup and cold-prefill requests. Complete streams do not mean correct or untruncated answers.

Request integrity audit
Complete SSE	Success counter delta	Prefix-hit token delta	Preemption delta	stop	length
268	268	0	0	177	91

Per-category batch details

56 formal batches and 7 ceiling_count batches; counting is excluded from the main table. The JSON download includes 261 individual request records.

Per-category batch details
Concurrency	Category	Throughput token/s	Stream decode token/s	Mean TTFT s	Output tokens	Batch duration s
C1	coding	65.59	71.56	0.267	200	3.049
C1	json	49.46	60.44	0.317	82	1.658
C1	narrative	26.88	28.16	0.267	138	5.133
C1	prose	35.52	37.74	0.246	133	3.744
C1	math	61.21	67.27	0.308	200	3.267
C1	reasoning	50.62	54.52	0.300	200	3.951
C1	summary	31.21	36.64	0.516	103	3.301
C1	format	61.41	75.61	0.324	102	1.661
C1	ceiling_count	78.19	84.21	0.230	239	3.057
C2	coding	115.92	64.71	0.291	400	3.451
C2	json	74.44	43.54	0.353	170	2.284
C2	narrative	43.14	23.90	0.335	285	6.606
C2	prose	47.66	27.15	0.320	263	5.518
C2	math	94.98	52.24	0.330	400	4.212
C2	reasoning	79.70	43.09	0.329	400	5.019
C2	summary	42.16	25.88	0.675	243	5.763
C2	format	115.69	72.03	0.367	208	1.798
C2	ceiling_count	147.29	80.07	0.271	478	3.245
C3	coding	160.94	59.68	0.350	600	3.728
C3	json	95.87	38.61	0.389	249	2.597
C3	narrative	51.09	18.35	0.301	440	8.611
C3	prose	60.60	23.36	0.291	389	6.419
C3	math	120.89	44.38	0.353	600	4.963
C3	reasoning	91.39	33.17	0.457	600	6.565
C3	summary	52.35	22.42	0.682	332	6.342
C3	format	141.95	59.26	0.384	310	2.184
C3	ceiling_count	204.49	73.50	0.267	717	3.506
C4	coding	201.49	56.88	0.373	800	3.970
C4	json	124.15	38.38	0.387	338	2.722
C4	narrative	58.82	15.87	0.326	557	9.470
C4	prose	71.59	20.43	0.316	506	7.068
C4	math	153.43	41.89	0.368	800	5.214
C4	reasoning	131.57	36.76	0.376	800	6.081
C4	summary	63.14	21.15	0.768	425	6.731
C4	format	132.69	42.97	0.402	435	3.278
C4	ceiling_count	236.04	63.74	0.285	956	4.050
C5	coding	234.28	53.24	0.371	1000	4.268
C5	json	133.21	32.86	0.398	422	3.168
C5	narrative	64.21	14.34	0.337	770	11.992
C5	prose	78.38	17.96	0.339	656	8.370
C5	math	192.32	42.43	0.403	1000	5.200
C5	reasoning	130.78	30.06	0.386	1000	7.647
C5	summary	66.83	17.88	0.880	566	8.469
C5	format	177.84	48.42	0.428	532	2.991
C5	ceiling_count	296.68	63.92	0.302	1195	4.028
C6	coding	255.94	48.66	0.413	1200	4.689
C6	json	148.71	30.55	0.407	506	3.403
C6	narrative	72.57	13.09	0.334	857	11.810
C6	prose	89.81	17.15	0.355	773	8.607
C6	math	203.20	37.10	0.400	1200	5.905
C6	reasoning	161.95	29.08	0.411	1200	7.410
C6	summary	75.29	16.46	1.156	706	9.377
C6	format	215.60	46.78	0.459	618	2.866
C6	ceiling_count	326.14	58.55	0.329	1434	4.397
C8	coding	308.79	43.66	0.503	1600	5.182
C8	json	178.63	27.30	0.465	668	3.740
C8	narrative	80.92	11.30	0.414	1155	14.274
C8	prose	105.89	15.35	0.349	996	9.406
C8	math	231.46	32.12	0.443	1600	6.913
C8	reasoning	167.87	23.66	0.771	1600	9.531
C8	summary	81.52	13.86	1.392	937	11.494
C8	format	218.18	37.29	0.504	853	3.910
C8	ceiling_count	406.55	54.39	0.324	1912	4.703

Workload and statistical methods

Tasks and runs: coding / json / narrative / prose / math / reasoning / summary / format; one full run; C8 is a local extension using the same method
Generation parameters: temperature=0; thinking=false; stream=true; usage tokens
Output limits: Default 200 tokens; summary 150; counting 256
Request composition: 268 = 232 formal + 29 counting + 3 warmup + 4 cold-prefill requests
Cold prompts and caches: Compilation caches reused; cold prompts do not imply cleared disk or compilation caches; no profiler during performance tests
Integrity and length limits: All 268 requests have complete SSE, DONE and usage; stop 177, length 91. Length-ended requests are retained, not counted as quality passes

Formal performance window: minimum host available 12.636 GiB, maximum host used 108.188 GiB. Final long-context window: 12.174 / 108.471 GiB. Host used includes system overhead, not just the model.
GB10 uses unified memory; CPU and GPU figures must not be added. Cgroup accounting misses some GPU-driver allocations, so roughly 20 GiB is not total model memory. The loading-log figure of about 81.69 GiB/rank is not a device peak.
Sampling at approximately one-second intervals captures observed extrema, not guaranteed instantaneous peaks. No Docker/cgroup OOM kill was observed in the listed windows; this is not a long-term stability guarantee.
No complete general answer-quality score was supplied. Functional passes, stream completeness, length termination and answer correctness are different measures.
Long-context evidence comprises two input lengths, one request each, with a middle-position needle. Full 300K, eight concurrent long requests, 1M, independent two-hour stability and fresh customer installation remain unverified.

C denotes concurrent request count. TTFT measures the wait for the first non-empty content output, not the full answer duration. Prefill and output decode rates are different metrics.

Plan your model deployment

Assess device configuration and integration for your tasks and concurrency needs.

Discuss this model deployment

DeepSeek-V4.1-Flash-1017

Model and validation scope

MegaCube deployment configuration

Verify weights and plan

Prepare nodes and interconnect

Check platform release coverage

Re-test actual workloads

MegaCube performance test results

Plan your model deployment