Qwen3.8 2.4T A95B
- Input
- $2.00/M tok
- Output
- $6.00/M tok
- Blended
- $3.00/M tok · 3:1
- Speed
- 38tok/s
- Context
- 262Ktokens
- Cost per task
- ~$2.5/taskest.
Open-weights checkpoint behind Qwen3.8 Max; priced at Qwen3.8 Max API rates.
Rankings
Coding
Mixed#54 / 137rank by quality56.6±6.6quality / 1003/7 benchmarks measuredWriting & chat
Mixed#28 / 162rank by quality70.6±7.3quality / 1002/5 benchmarks measuredHard reasoning
Verified#29 / 218rank by quality74.0±3.7quality / 1003/4 benchmarks measuredAgents
Verified#17 / 187rank by quality80.6±5.9quality / 1003/7 benchmarks measured
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
- 24× B200 180GB
- FP8 weights
- Throughput
- ~27,655 tok/s
- many requests batched
- Price
- $0.66–5.85 /M tok
- busy → light use
- On your own machine
- Too large
- needs data-center GPUs
Assumptions (5)
- FP8 weights (2400B params, 95B active per token) + 25% KV-cache headroom ≈ 3000 GB
- 24× B200 180GB at $5.98–$14.24/GPU-hour on-demand (2026-09-24)
- ~27,655 output tok/s aggregate at batch 931 (bandwidth-bound); MoE compute scales with active params
- Blended 3:1 input:output; prefill ~228,292 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)~82.4%
- #1Claude Opus 596.0%
- #2Claude Sonnet 585.2%
- ~#3Qwen3.8 2.4T A95BEstimated82.4%
- #3Beam 1.0 501B A23B81.6%
No published score — estimated from related benchmarks
SWE-bench Pro (Public)Estimated
~#9 of 24 (estimated)~64.1%
- #1Claude Fable 5.293.0%
- #2Claude Opus 5.589.9%
- #8Hy4 preview65.7%
- ~#9Qwen3.8 2.4T A95BEstimated64.1%
- #9GPT-5.6 Terra63.4%
No published score — estimated from related benchmarks
DeepSWE v1.1Lab-reported
#35 of 4556.6%
- #1Gemini 4 Argon78.8%
- #2DeepSeek V4.1 Flash 101777.0%
- #34DeepSeek V4 Pro (0813)57.2%
- #35Qwen3.8 2.4T A95B56.6%
- #36Nex-N2.5-Pro55.8%
huggingface.co · 2026-08-12
Terminal-Bench 4.0Verified
#39 of 10111.1%
- #1Kimi K3.1 Pro67.2%
- #2Claude Sonnet 5.563.6%
- #36Kimi K312.6%
- #39Qwen3.8 2.4T A95B11.1%
- #40Muse Spark Open8.1%
artificialanalysis.ai · 2026-09-24
LiveCodeBenchEstimated
~#10 of 28 (estimated)~82.9%
- #1Qwen3.8-Omni-Flash92.6%
- #2Qwen3.8 Flash-Next91.9%
- #9Nemotron 3.5 Super VL NVFP487.4%
- ~#10Qwen3.8 2.4T A95BEstimated82.9%
- #10Ternary Bonsai 27B82.8%
No published score — estimated from related benchmarks
SciCodeVerified
#32 of 11854.1%
- #1Claude Fable 5.267.7%
- #2Claude Opus 5.566.9%
- #32Ling 3.1 Flash54.1%
- #32Qwen3.8 2.4T A95B54.1%
- #34DeepSeek-V4-Pro-0813-nvfp4-DSpark53.8%
artificialanalysis.ai · 2026-09-24
AA Coding Agent IndexEstimated
~#21 of 26 (estimated)~40.0
- #1Claude Opus 5.561.7
- #2Claude Opus 559.7
- #20Kimi K341.4
- ~#21Qwen3.8 2.4T A95BEstimated40.0
- #21Qwen3.8 Max (0902)39.2
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
LMArena Text, Non-English (Elo)Estimated
~#21 of 57 (estimated)~1447 Elo
- #1Gemini 4 ArgonFew votes1512
- #2Claude Opus 5.51497
- #19Grok 4.61447
- ~#21Qwen3.8 2.4T A95BEstimated1447
- #21Gemini 3.5 Flash-Lite1445
No published score — estimated from related benchmarks
LMArena Text (Elo)Estimated
~#20 of 57 (estimated)~1459 Elo
- #1Gemini 4 Argon1525
- #2Claude Opus 5.51507
- #19Claude Sonnet 51461
- ~#20Qwen3.8 2.4T A95BEstimated1459
- #20GPT-6 Sol1457
No published score — estimated from related benchmarks
#12 of 251843 Elo
eqbench.com · 2026-09-24
Hard reasoning
GPQA DiamondVerified
#5 of 19293.5%
- #1GPT-6 Astra96.1%
- #2Gemini 3.8 Flash95.3%
- #5Muse Spark 1.393.5%
- #5Qwen3.8 2.4T A95B93.5%
- #5Step 5 Preview93.5%
artificialanalysis.ai · 2026-09-24
Humanity's Last ExamVerified
#30 of 19842.4%
- #1Claude Fable 5.274.3%
- #2Claude Opus 5.561.4%
- #29GPT-5.3 Codex42.5%
- #30Qwen3.8 2.4T A95B42.4%
- #31GLM-5.342.3%
artificialanalysis.ai · 2026-09-24
AIME (latest)Estimated
~#3 of 34 (estimated)~97.4%
No published score — estimated from related benchmarks
Agents
τ²-bench (Telecom)Estimated
~#3 of 120 (estimated)~96.7%
- #1Step 3.7 Flash98.5%
- #2DeepSeek-V4-Pro-0813-nvfp4-DSpark98.3%
- ~#3Qwen3.8 2.4T A95BEstimated96.7%
- #3Gemini 3.1 Pro95.6%
No published score — estimated from related benchmarks
OSWorld-VerifiedEstimated
~#3 of 15 (estimated)~82.6%
No published score — estimated from related benchmarks
OSWorld 2.0Estimated
~#3 of 5 (estimated)~57.2%
No published score — estimated from related benchmarks
BrowseCompEstimated
~#6 of 22 (estimated)~89.4%
- #1Atria Dawn Preview92.5%
- #2GPT-6 Astra91.5%
- #5Nex-N2.5-Pro89.7%
- ~#6Qwen3.8 2.4T A95BEstimated89.4%
- #6Step 5 Preview88.7%
No published score — estimated from related benchmarks
GDPval-AA (Elo)Verified
#16 of 911597 Elo
- #1Claude Opus 5.51866
- #2Muse Wonder Full1857
- #15DeepSeek V4.1 Flash1600
- #16Qwen3.8 2.4T A95B1597
- #17GPT-6.1 Sol1592
artificialanalysis.ai · 2026-09-24
AutomationBench-AAVerified
#20 of 8457.2%
- #1Claude Fable 5.276.4%
- #2Muse Wonder Full71.3%
- #19Muse Spark 1.357.9%
- #20Qwen3.8 2.4T A95B57.2%
- #21DeepSeek V4 Pro (0813)56.7%
artificialanalysis.ai · 2026-09-24