Qwen3.8 Flash-Next
- Input
- $0.15/M tok
- Output
- $0.47/M tok
- Blended
- $0.23/M tok · 3:1
- Speed
- 54tok/s
- Context
- 262Ktokens
- Cost per task
- ~$0.20/taskest.
Open weights (125B total / 6B active); price is the hosted Qwen3.8-Flash.
Rankings
Coding
Mixed#34 / 137rank by quality64.3±5.3quality / 1004/7 benchmarks measuredWriting & chat
Lab-reported#12 / 162rank by quality77.6±9.7quality / 1001/5 benchmarks measuredPrice frontierVision
Mixed#22 / 98rank by quality72.2±5.2quality / 1003/6 benchmarks measuredHard reasoning
Verified#36 / 218rank by quality71.9±3.7quality / 1003/4 benchmarks measuredAgents
Verified#18 / 187rank by quality79.6±5.9quality / 1003/7 benchmarks measuredPrice frontier
Self-hostingEstimated
How we estimate →What it would cost to run these open weights yourself on rented GPUs, compared with the API price above.
- Hardware
- 2× H200 141GB
- FP8 weights
- Throughput
- ~8,910 tok/s
- many requests batched
- Price
- $0.087–0.72 /M tok
- busy → light use
- On your own machine
- Mac Studio M5 Max 128GB (int4, ~140 tok/s single-stream)
- single consumer GPU or Mac
Assumptions (5)
- FP8 weights (125B params, 6B active per token) + 25% KV-cache headroom ≈ 156 GB
- 2× H200 141GB at $3.59–$7.91/GPU-hour on-demand (2026-09-24)
- ~8,910 output tok/s aggregate at batch 320 (bandwidth-bound); MoE compute scales with active params
- Blended 3:1 input:output; prefill ~155,846 tok/s
- Low = 75% utilization at the low GPU price; high = 20% at the high price
Benchmark scores
~ italic, dashed = no published score yet, estimated from related benchmarks.
Coding
SWE-bench VerifiedEstimated
~#3 of 21 (estimated)~83.0%
- #1Claude Opus 596.0%
- #2Claude Sonnet 585.2%
- ~#3Qwen3.8 Flash-NextEstimated83.0%
- #3Beam 1.0 501B A23B81.6%
No published score — estimated from related benchmarks
SWE-bench Pro (Public)Estimated
~#8 of 24 (estimated)~65.9%
- #1Claude Fable 5.293.0%
- #2Claude Opus 5.589.9%
- #7Beam 1.0 501B A23B66.1%
- ~#8Qwen3.8 Flash-NextEstimated65.9%
- #8Hy4 preview65.7%
No published score — estimated from related benchmarks
DeepSWE v1.1Lab-reported
#32 of 4558.7%
- #1Gemini 4 Argon78.8%
- #2DeepSeek V4.1 Flash 101777.0%
- #31Muse Spark Open60.0%
- #32Qwen3.8 Flash-Next58.7%
- #33Qwen3.8-Omni-Flash57.8%
huggingface.co · 2026-08-26
Terminal-Bench 4.0Verified
#28 of 10125.3%
- #1Kimi K3.1 Pro67.2%
- #2Claude Sonnet 5.563.6%
- #26MiniMax-M3.1-Flash25.8%
- #28Qwen3.8 Flash-Next25.3%
- #29MiMo-V2.6-Flash22.7%
artificialanalysis.ai · 2026-09-24
SciCodeVerified
#44 of 11850.6%
- #1Claude Fable 5.267.7%
- #2Claude Opus 5.566.9%
- #43DeepSeek V4 Pro (0813)51.0%
- #44Qwen3.8 Flash-Next50.6%
- #45Inkling Small49.7%
artificialanalysis.ai · 2026-09-24
AA Coding Agent IndexEstimated
~#20 of 26 (estimated)~43.0
- #1Claude Opus 5.561.7
- #2Claude Opus 559.7
- #18Step 5 Preview43.1
- ~#20Qwen3.8 Flash-NextEstimated43.0
- #20Kimi K341.4
No published score — estimated from related benchmarks
Not comparable across labs (1)
Kept for reference, not counted in rankings: these use a lab-specific task set, answer key or scoring, so scores can't be compared fairly across labs. Only published scores are shown.
Writing & chat
IFBenchLab-reported
#5 of 13981.3%
- #1MiniMax-M382.9%
- #2Qwen3.8 2.4T A95B82.8%
- #4Nemotron 3 Ultra81.4%
- #5Qwen3.8 Flash-Next81.3%
- #6Qwen3.8 Flash-Next NVFP481.0%
huggingface.co · 2026-08-26
LMArena Text, Non-English (Elo)Estimated
~#18 of 57 (estimated)~1456 Elo
- #1Gemini 4 ArgonFew votes1512
- #2Claude Opus 5.51497
- #17GPT-5.6 Terra1456
- ~#18Qwen3.8 Flash-NextEstimated1456
- #18DeepSeek V4 Pro (0813)1448
No published score — estimated from related benchmarks
LMArena Text (Elo)Estimated
~#17 of 57 (estimated)~1468 Elo
- #1Gemini 4 Argon1525
- #2Claude Opus 5.51507
- #16DeepSeek V4.1 Flash1474
- ~#17Qwen3.8 Flash-NextEstimated1468
- #17GPT-5.6 Terra1466
No published score — estimated from related benchmarks
EQ-Bench Creative Writing v3 (Elo)Estimated
~#9 of 26 (estimated)~1952 Elo
- #1GPT-6 Astra2173
- #2Claude Fable 5.12162
- #8Grok 4.72007
- ~#9Qwen3.8 Flash-NextEstimated1952
- #9Muse Spark 1.31906
No published score — estimated from related benchmarks
Vision
OmniDocBench v1.5Estimated
~#4 of 8 (estimated)~88.9%
- #1MiniMax-M391.6%
- #2Kimi K391.1%
- #2Qwen3.8 27B91.1%
- ~#4Qwen3.8 Flash-NextEstimated88.9%
- #4Gemini 3.1 Pro85.3%
No published score — estimated from related benchmarks
GDP.pdf (AA)Verified
#23 of 4715.6%
- #1Muse Wonder Full32.6%
- #2GPT-6 Astra31.0%
- #22Qwen3.8 27B16.6%
- #23Qwen3.8 Flash-Next15.6%
- #24GLM-5.3 Flash15.4%
artificialanalysis.ai · 2026-09-24
ChartographyEstimated
~#10 of 25 (estimated)~29.4%
- #1Claude Fable 5.273.4%
- #2GPT-6 Astra71.0%
- #9GPT-5.531.0%
- ~#10Qwen3.8 Flash-NextEstimated29.4%
- #10GPT-6 Luna29.1%
No published score — estimated from related benchmarks
Hard reasoning
GPQA DiamondVerified
#17 of 19292.3%
- #1GPT-6 Astra96.1%
- #2Gemini 3.8 Flash95.3%
- #17Hy4 preview92.3%
- #17Qwen3.8 Flash-Next92.3%
- #19GLM-5.3 Flash NVFP492.1%
artificialanalysis.ai · 2026-09-24
Humanity's Last ExamVerified
#40 of 19838.0%
- #1Claude Fable 5.274.3%
- #2Claude Opus 5.561.4%
- #39Grok Build 0.1 061638.3%
- #40Qwen3.8 Flash-Next38.0%
- #41Motif 337.0%
artificialanalysis.ai · 2026-09-24
AIME (latest)Estimated
~#2 of 34 (estimated)~99.1%
No published score — estimated from related benchmarks
Agents
τ²-bench (Telecom)Estimated
~#2 of 120 (estimated)~98.3%
No published score — estimated from related benchmarks
OSWorld-VerifiedEstimated
~#3 of 15 (estimated)~83.1%
No published score — estimated from related benchmarks
OSWorld 2.0Estimated
~#3 of 5 (estimated)~56.5%
No published score — estimated from related benchmarks
BrowseCompEstimated
~#5 of 22 (estimated)~89.9%
- #1Atria Dawn Preview92.5%
- #2GPT-6 Astra91.5%
- #4Claude Opus 590.8%
- ~#5Qwen3.8 Flash-NextEstimated89.9%
- #5Nex-N2.5-Pro89.7%
No published score — estimated from related benchmarks
GDPval-AA (Elo)Verified
#14 of 911612 Elo
- #1Claude Opus 5.51866
- #2Muse Wonder Full1857
- #13Grok 4.61632
- #14Qwen3.8 Flash-Next1612
- #15DeepSeek V4.1 Flash1600
artificialanalysis.ai · 2026-09-24
Not comparable across labs (1)
Kept for reference, not counted in rankings: these use a lab-specific task set, answer key or scoring, so scores can't be compared fairly across labs. Only published scores are shown.