Skip to main content
Alibaba logo
QwenVision-language · Open weights

Qwen3.8-27B

Dense vision-language reasoning from Qwen: 27B language parameters, a native 262K-token context, and text plus image input.

On 1 × NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB), Qwen3.8-27B measured 26.12 tokens/s at concurrency 1 and 179.18 tokens/s aggregate at concurrency 8 in a 10-minute sustained test (1,024 input / 256 output tokens, measured 2026-10-14).

Context262,144Params27BI/OText + images → textApache 2.0Measured on RTX PRO 6000
View benchmarksModel card

From the model card

Specifications

Published by the model author. Reviewed 2026-10-11.

Publisher model card
Architecture
Dense vision-language model
Language parameters
27B
Layers / hidden size
64 / 5,120
Published native context
262,144 tokens
Extended context
Up to 1M with the documented configuration
Input → output
Text + images → text
License
Apache 2.0

Measured by EcoHash

How it scales on one RTX PRO 6000

Closed-loop sustained load with 256 output tokens fixed. Concurrency is the level tested, not a capacity limit.

1 × NVIDIA RTX PRO 6000 Blackwell Server EditionvLLM 0.29.0BF16Atlanta, Georgia, US
Aggregate output throughputtokens/s vs concurrent requests
Aggregate output throughput. 1,024 in / 256 out: 1 → 26 tokens/s, 2 → 49 tokens/s, 4 → 93 tokens/s, 8 → 179 tokens/s, 16 → 309 tokens/s. 8,192 in / 256 out: 1 → 23 tokens/s, 2 → 40 tokens/s, 4 → 65 tokens/s, 8 → 98 tokens/s01002003004001248161,024 in / 256 out8,192 in / 256 out
1,024 in / 256 out8,192 in / 256 outx axis: concurrent requests (log₂)
Time per output token, P50ms vs concurrent requests
Time per output token, P50. 1,024 in / 256 out: 1 → 37.79 ms, 2 → 39.38 ms, 4 → 40.04 ms, 8 → 39.94 ms, 16 → 44.57 ms. 8,192 in / 256 out: 1 → 38.08 ms, 2 → 44.67 ms, 4 → 45.91 ms, 8 → 63.51 ms0204060801248161,024 in / 256 out8,192 in / 256 out
1,024 in / 256 out8,192 in / 256 outsustained load · 2026-10-14
1,024-token prompts: 11.8× the single-request throughput at 16 concurrent. Aggregate output rose from 26 to 309 tokens/s between concurrency 1 and 16, while median time per output token grew 1.2× and P95 time to first token reached 2.46 s.
8,192-token prompts: 4.2× the single-request throughput at 8 concurrent. Aggregate output rose from 23 to 98 tokens/s between concurrency 1 and 8, while median time per output token grew 1.7× and P95 time to first token reached 7.58 s.
Stable for 10 minutes at 8 concurrent. A 606-second soak (424 requests) measured 179 tokens/s, within 0.1% of the 126-second run at the same level, with 0 failed requests.

Test results at a glance

Pick the column closest to your workload. Each column pairs throughput and latency from one run; higher total throughput trades off latency.

Metric
Single request
1,024 in / 256 out
concurrency 1
Sustained concurrent load
1,024 in / 256 out
concurrency 8 · 10-min soak
Long input · single request
8,192 in / 256 out
concurrency 1
Long input · concurrent load
8,192 in / 256 out
concurrency 4
met recorded internal targets
Output throughput(tokens/s)26.12179.1823.3165.41
TTFT P50 / P95(ms)161.84 / 166.691,243.27 / 1,249.31,266.58 / 1,290.233,291.7 / 3,804.32
TPOT P50 / P95(ms per token)37.79 / 37.7939.94 / 42.1138.08 / 38.0945.91 / 53.41
End-to-end P50 / P95(s)9.8 / 9.811.43 / 11.4310.98 / 1115.5 / 15.51
Request success rate100%100%100%100%
Measured requests644246467
Test duration(min)10.510.111.74.4

TTFT is the wait for the first output token. TPOT is the average time per subsequent token. P50 / P95 are request percentiles. Recorded concurrency is the level tested, not a service SLA or a physical capacity limit. Rounds are reported separately, never averaged. “Met recorded internal targets” repeats the pass recorded with that run; it is not a service commitment.

All 14 measured rows1,024 and 8,192-token inputs · fixed rounds and sustained load · 1,231 requests

Measured workload profiles

Each row pairs throughput and latency from one run. Aggregate throughput counts output tokens from all requests, including time spent processing the input. P50 / P95 are request percentiles; TPOT is the average time per output token after the first.

Tokens in / outConcurrentRunOutput tokens/sReq/sTTFT P50 / P95 (ms)TPOT P50 / P95 (ms/token)End-to-end P50 / P95 (s)Requests
1,024 / 25612026-10-14 · Sustained load · 627 s soak
26.12
0.1161.84 / 166.6937.79 / 37.799.8 / 9.864
1,024 / 25622026-10-14 · Sustained load
48.69
0.19291.37 / 331.639.38 / 39.5910.37 / 10.3765
1,024 / 25642026-10-14 · Sustained load
93.2
0.36627.13 / 629.5240.04 / 40.8910.84 / 10.8467
1,024 / 25682026-10-14 · Sustained load
178.98
0.71,256.36 / 1,268.6939.94 / 42.1311.44 / 11.4588
1,024 / 25682026-10-14 · Sustained load · 606 s soak
179.18
0.71,243.27 / 1,249.339.94 / 42.1111.43 / 11.43424
1,024 / 256162026-10-14 · Sustained load
309.13
1.211,877.75 / 2,459.6744.57 / 48.9613.25 / 13.32160
1,024 / 25612026-10-11 · Fixed rounds
26.18
0.1162.95 / 163.6137.71 / 37.719.78 / 9.7824
1,024 / 25642026-10-11 · Fixed rounds
94.79
0.37625.77 / 632.4539.9 / 40.7410.8 / 10.8124
Reliability · lowest success rate 100% · failures 0 · 1,231 measured requests across 14 rowsFixed-length outputs: P50 and P95 can be nearly equal. Not representative of variable-length requests.

Measured by EcoHash

Memory & storage

Resource values use the saved test configuration (2026-10-11). Workload rows have their own dates and sequence limits. These are recorded values, not minimum requirements.

Fits on one 96 GB GPU, no CPU offloading88.7 GB peak of 96 GB

88.7 GB of GPU memory at peak on a 96 GB GPU, including the KV cache reserved by the engine. 7.3 GB was not used at peak.

GPU memory · peak workload
88.7 GB
Including KV cache reserved by the engine
Host RAM · peak
7.3 GB
Process memory, excludes page cache
Weight storage
55.6 GB
On-disk storage for the served checkpoint
Cache footprint (downloads + runtime)
0.1 GB
Separate from the served weights

Host RAM is the summed process memory and excludes the filesystem page cache. Cache footprint is separate from the served weights and can include downloaded checkpoints that were not served.

Source & use

License & links

Where this model comes from and what you may do with it.

Apache 2.0Commercial use permitted

Commercial use is permitted under Apache 2.0, subject to its notice and other requirements. Review the license for the exact checkpoint you deploy.

Publisher
Qwen
Model ID
qwen3.8-27b
License
Apache 2.0
Model card / source
Qwen on Hugging Face
Weight revision tested
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0

Specifications reviewed 2026-10-11. Measurements are EcoHash’s own, recorded with their full methodology below.

Measured by EcoHash

Methodology & provenance

How each number on this page was produced, with the raw archives that back it.

Run provenance2 campaigns · 3 raw archives
Sustained load2026-10-14

10 rows · 1,135 measured requests · server sequence limit 64. Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL.

SHA256 ffd3b68b4e257230c1232a0e1fdd89ebf4081a28fbb9c58765a32ae67565f4ac · f2a08777c64d2b00876a0110d4a07c9a83db337bdcac37b2acdad2ce6dae7611

Fixed rounds2026-10-11

4 rows · 96 measured requests · server sequence limit 4. Local streaming HTTP, native token lengths; 3 rounds × 8 requests and 4 warmups per profile. Throughput uses summed measured round durations. Fixed 256-token outputs with ignore_eos=true; prefix caching off.

SHA256 d41a9b6055b6e90512b9932915a65005fda70deb7186c60f7948c8b9e8b5ac64

Every row’s recorded conditions
  • 1,024 in / 256 out · concurrency 1

    2026-10-11 · Fixed rounds · 234.7 measured seconds · server sequence limit 4

    Local streaming HTTP, native token lengths; 3 rounds × 8 requests and 4 warmups per profile. Throughput uses summed measured round durations. Fixed 256-token outputs with ignore_eos=true; prefix caching off. This run did not establish a maximum serving capacity. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: d41a9b6055b6e90512b9932915a65005fda70deb7186c60f7948c8b9e8b5ac64

  • 1,024 in / 256 out · concurrency 4

    2026-10-11 · Fixed rounds · 64.8 measured seconds · server sequence limit 4

    Local streaming HTTP, native token lengths; 3 rounds × 8 requests and 4 warmups per profile. Throughput uses summed measured round durations. Fixed 256-token outputs with ignore_eos=true; prefix caching off. This run did not establish a maximum serving capacity. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: d41a9b6055b6e90512b9932915a65005fda70deb7186c60f7948c8b9e8b5ac64

  • 8,192 in / 256 out · concurrency 1

    2026-10-11 · Fixed rounds · 262.9 measured seconds · server sequence limit 4

    Local streaming HTTP, native token lengths; 3 rounds × 8 requests and 4 warmups per profile. Throughput uses summed measured round durations. Fixed 256-token outputs with ignore_eos=true; prefix caching off. This run did not establish a maximum serving capacity. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: d41a9b6055b6e90512b9932915a65005fda70deb7186c60f7948c8b9e8b5ac64

  • 8,192 in / 256 out · concurrency 4

    2026-10-11 · Fixed rounds · 92.9 measured seconds · server sequence limit 4

    Local streaming HTTP, native token lengths; 3 rounds × 8 requests and 4 warmups per profile. Throughput uses summed measured round durations. Fixed 256-token outputs with ignore_eos=true; prefix caching off. This run did not establish a maximum serving capacity. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: d41a9b6055b6e90512b9932915a65005fda70deb7186c60f7948c8b9e8b5ac64

  • 1,024 in / 256 out · concurrency 1

    2026-10-14 · Sustained load · 627.1 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 2 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 25 s; success ≥ 99%. This profile met those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: ffd3b68b4e257230c1232a0e1fdd89ebf4081a28fbb9c58765a32ae67565f4ac

  • 1,024 in / 256 out · concurrency 2

    2026-10-14 · Sustained load · 341.8 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 2 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 25 s; success ≥ 99%. This profile met those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: ffd3b68b4e257230c1232a0e1fdd89ebf4081a28fbb9c58765a32ae67565f4ac

  • 1,024 in / 256 out · concurrency 4

    2026-10-14 · Sustained load · 184.0 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 2 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 25 s; success ≥ 99%. This profile met those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: ffd3b68b4e257230c1232a0e1fdd89ebf4081a28fbb9c58765a32ae67565f4ac

  • 1,024 in / 256 out · concurrency 8

    2026-10-14 · Sustained load · 125.9 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 2 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 25 s; success ≥ 99%. This profile met those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: ffd3b68b4e257230c1232a0e1fdd89ebf4081a28fbb9c58765a32ae67565f4ac

  • 1,024 in / 256 out · concurrency 16

    2026-10-14 · Sustained load · 132.5 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 2 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 25 s; success ≥ 99%. This profile did not meet those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: ffd3b68b4e257230c1232a0e1fdd89ebf4081a28fbb9c58765a32ae67565f4ac

  • 8,192 in / 256 out · concurrency 1

    2026-10-14 · Sustained load · 702.8 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 5 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 40 s; success ≥ 99%. This profile met those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: ffd3b68b4e257230c1232a0e1fdd89ebf4081a28fbb9c58765a32ae67565f4ac

  • 8,192 in / 256 out · concurrency 8

    2026-10-14 · Sustained load · 184.8 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 5 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 40 s; success ≥ 99%. This profile did not meet those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: ffd3b68b4e257230c1232a0e1fdd89ebf4081a28fbb9c58765a32ae67565f4ac

  • 1,024 in / 256 out · concurrency 8

    2026-10-14 · Sustained load · 605.8 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 2 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 25 s; success ≥ 99%. This profile met those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: ffd3b68b4e257230c1232a0e1fdd89ebf4081a28fbb9c58765a32ae67565f4ac

  • 8,192 in / 256 out · concurrency 2

    2026-10-14 · Sustained load · 418.9 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 5 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 40 s; success ≥ 99%. This profile met those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: f2a08777c64d2b00876a0110d4a07c9a83db337bdcac37b2acdad2ce6dae7611

  • 8,192 in / 256 out · concurrency 4

    2026-10-14 · Sustained load · 262.2 measured seconds · server sequence limit 64

    Synthetic native-token inputs; closed-loop client. Internal provisional latency targets. Inter-chunk gaps are not token-level ITL. OOM requires server-log review. Internal comparison targets: TTFT P95 ≤ 5 s; TPOT P95 ≤ 80 ms; end-to-end P95 ≤ 40 s; success ≥ 99%. This profile met those targets. These are not a service SLA. Measurement: local HTTP, excluding public internet and gateway latency. TTFT measures request send to first non-empty output; TPOT = (last output time - first output time) / (actual output tokens - 1). P50/P95 are computed across individual successful requests in this profile; warmups excluded.

    Raw archive SHA256: f2a08777c64d2b00876a0110d4a07c9a83db337bdcac37b2acdad2ce6dae7611

Recorded test configuration & full methodologyrevision 5 · updated 2026-10-14
Hardware SKU
1 × NVIDIA RTX PRO 6000 Blackwell Server Edition; 96 GB VRAM (97,887 MiB reported)
Deployment region
Atlanta, Georgia, US
GPU power limit
600 W
NVIDIA driver
580.105.08
CUDA version
13.0
Engine & version
vLLM 0.29.0; PyTorch 2.13.0+cu130; Transformers 5.16.1
Container image digest
vllm/vllm-openai@sha256:082ca6f035279109041ffd3fe0695cb568b29bc580b35c4f297a66a08b216c1b
Weight revision
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
Precision / quantization
BF16; no quantization
CPU offloading
None
Test date
2026-10-11
Warmup runs
4 per profile (16 total)
Measured repeats
3 rounds × 8 requests per profile
Sample count
24 measured requests per profile; 96 across four profiles; plus 16 warmups and 1 smoke check
Input tokens
1,024 baseline; 8,192 secondary profile
Output tokens
256 fixed
Reasoning setting
Synthetic completions; functional chat smoke: enable_thinking=false
Test context limit (tokens)
9,216
KV cache
gpu_memory_utilization=0.85; prefix caching off
Tested concurrency levels
1 and 4
Seed
20260920
Test notes
Workload: synthetic native-token inputs of 1,024 or 8,192 tokens; 256 fixed output tokens; temperature 0; ignore_eos=true. The workload table contains all four 2026-10-11 profiles (96 measured requests) and all 2026-10-14 capacity profiles (1135 measured requests). Local HTTP; public internet and gateway latency excluded. Each row records its own concurrency, sample count, duration and server sequence limit. The original flat performance and resource measurements below refer to 2026-10-11; the newer latency and throughput results are in the workload table. Latency definitions: TTFT is request send to first non-empty output. TPOT is (last output time - first output time) / (actual output tokens - 1), excluding first-token wait. P50/P95 are percentiles across individual successful requests within each profile; warmups excluded. Chunk intervals are not exact per-token ITL. Capacity: 2026-10-14 used server max_num_seqs=64, with closed-loop concurrency scans and a 10-minute short-input soak. Internal latency targets and passing status are recorded per profile; these are not a service SLA or a physical maximum. Request failures were zero; startup allocation OOM warnings recovered before service readiness and are not counted as failed inference requests. Quality: not evaluated. One functional chat smoke check passed (expected answer 323); its reasoning setting applies to that check. GPU memory: workload peak 84,572 MiB = 88.6802 GB (671 samples inside the workload window, startup and uncertain boundary samples excluded), includes KV-cache reservation. Full-lifecycle peak, including initialization and profiling, was the same. Host RAM: peak summed process PSS 7.2578 GB (benchmark client + server), excludes filesystem page cache; full-lifecycle PSS peak 7.6737 GB. Container cgroup peak including page cache: 103.0775 GB. 866 telemetry samples, max interval 1.352 s. Storage: served shard set 55.5630 GB (model.safetensors.index.json); total HF cache 55.5860 GB; runtime cache 0.1205 GB. Remaining HF cache plus runtime cache: 0.1435 GB. 1 GB = 1e9 bytes. Not measured separately: idle VRAM, cold and warm start, output storage. Startup to readiness 215.05 s includes download and compile; it is not an isolated cold-start measurement. Raw archive SHA256: d41a9b6055b6e90512b9932915a65005fda70deb7186c60f7948c8b9e8b5ac64

Updated 2026-10-14 · revision 5

Choose Qwen3.8-27B for text and image workflows

Explore Qwen3.8-27B text and image use cases, RTX Pro 6000 benchmarks and deployment guidance. Check model configuration and service availability.

When to choose it

  • Evaluating text reasoning and image-assisted tasks such as answering questions about supplied screenshots or documents.
  • Inspecting a measured single-GPU serving configuration before running your own workload on RTX Pro 6000.

Before you integrate

  • Published model context, the endpoint's configured context and the input lengths in a benchmark are different limits. Check each against your workload.
  • Benchmark speed depends on precision, input length, output length and concurrency. Compare matching test conditions rather than treating one result as a guaranteed rate.
  • Review the deployment guide and live quote before launching. A published template does not establish shared API availability.
Read the publisher's model card →

Guides and next steps

More model profiles

Specifications & test evidence

All models
OpenAI · Languagegpt-oss-120bSpecifications, measurements & methodologyAutoTrust AI · LanguageJEV-27BSpecifications, measurements & methodologyLightricks · Video + audioLTX-2.5Specifications, real outputs & license