Skip to content
ModelCap

Qwen3.8 27B

Arena #112Official publisher releaseOpen weights6.9d #70 → #68

Qwen·#68 of 371 on the current board·cheaper than #65 MiMo-V2.5

Qwen3.8 27B is a language model from Qwen. It ranks #68 of 371 on ModelCap's current board, with an Index score of 63.1, measured on 4 public benchmarks. Its score range of 58.5–67.7 spans positions #51–#92. It is listed at $0.06 per million input tokens and $2.20 per million output tokens across 24 providers. It reads up to 1M tokens of context, and its weights are openly downloadable.

Figures as of 19 Nov 2026, 05:16 UTC · download the public dataset · methodology

ModelCap Index · Current board
63.1#68
Measured on 4 public benchmarks · range 58.5–67.7
Evidence as of 18 Nov 2026, 14:36 UTC
Market Gravity
62.8Decreased by 1.0
Secondary market signal
Input
$0.06
per 1M tokens
Output
$2.20
per 1M tokens
Context
1M
tokens
Providers
24
endpoints

Providers

Uptime measured over the last 30 minutes
ProviderInputOutputContextQuantUptime
StreamLakeCheapest$0.175$1.051M—100.0%
Chutesfp8$0.039$1.25262Kfp8100.0%
Near AIfp8$0.04$1.35262Kfp8100.0%
Cerebrasfp16$0.99$1.4966Kfp16100.0%
Darkbloomfp4$0.035$1.50262Kfp4100.0%
Mancer 2fp8$0.125$1.50262Kfp8100.0%
DeepInfrafp8$0.15$1.88262Kfp8100.0%
Ionstreamfp8$0.04$1.95262Kfp8100.0%
AkashMLfp8$0.225$1.98262Kfp8100.0%
Parasailfp8$0.24$2.20262Kfp8100.0%
Waferfp8$0.06$2.20262Kfp8100.0%
Modal$0.04$2.30262K—100.0%
DekaLLM$0.04$2.50262K—100.0%
ModelRunfp4$0.20$2.50262Kfp4100.0%
Rekafp8$0.15$2.50262Kfp8100.0%
Alibaba$0.425$2.551M—100.0%
SiliconFlowfp8$0.425$2.55262Kfp8—
CoreWeavefp8$0.40$3.00262Kfp8100.0%
Nebiusfp4$0.45$3.001Mfp4100.0%
Novita$0.42$3.001M—100.0%
Cloudflare$0.45$3.20262K——
GMICloudfp8$0.45$3.201Mfp8—
PrimeIntellectfp8$0.45$3.20262Kfp8100.0%
Venicefp8$0.45$3.20262Kfp8100.0%

Hugging Face Inference Providers

Exact repository identity

Provider economics and telemetry published by the Hugging Face Router. These are operational signals, not benchmark results.

ProviderInputOutputContextTTFTThroughputFeatures
near-ai
live
$0.04$1.35262K1413 ms83.0 tok/stools · structured
cerebras
live
$0.99$1.4966K270 ms560.0 tok/stools
mancer
live
$0.125$1.50262K1409 ms58.0 tok/stools · structured
deepinfra
live
$0.15$1.88262K781 ms39.0 tok/stools · structured
ionstream
live
$0.04$1.95262K1003 ms67.0 tok/stools · structured
akashml
live
$0.225$1.98262K992 ms33.0 tok/stools · structured
parasail
live
$0.24$2.20262K540 ms74.0 tok/stools · structured
wafer
live
$0.06$2.20262K472 ms134.0 tok/stools · structured
modal
live
$0.04$2.30262K378 ms103.0 tok/stools · structured
dekallm
live
$0.04$2.50262K741 ms117.0 tok/stools · structured
modelrun
live
$0.20$2.50262K258 ms178.0 tok/stools · structured
nscale
live
$0.20$2.501M453 ms106.7 tok/stools · structured
reka
live
$0.15$2.50262K1244 ms67.0 tok/stools · structured
siliconflow
live
$0.425$2.55262K1092 ms49.5 tok/stools
coreweave
live
$0.40$3.00262K467 ms107.0 tok/stools · structured
nebius
live
$0.45$3.001M946 ms197.0 tok/stools · structured
novita
live
$0.42$3.001M2225 ms39.0 tok/stools
ovhcloud
live
$0.47$3.19262K874 ms76.0 tok/stools
cloudflare
live
$0.45$3.20262K322 ms46.5 tok/stools · structured
primeintellect
live
$0.45$3.20262K294 ms129.0 tok/stools · structured
venice
live
$0.45$3.20262K1158 ms47.5 tok/stools · structured
featherless-ai
live
——————

Reasoning efforts

Best supported result can inform ModelCap

Published Max, X-High, High and other configurations for this model. These are sourced scores, not head-to-head essays.

EffortAgent scoreSource rankSampleSource
Default
Qwen3.8-27B
-5.50%
CI -6.4–-4.6%
#518.5M obs
64K sessions
LMArena Agent
Match 100%

Reported evaluation leads (143)

Not independent measurement

Structured rows reported in this exact model repository at the pinned revision below. They are the publisher talking about itself: never measured evidence, never a score input on their own. They remain quarantined until dataset revision, harness, configuration and identity pass ModelCap’s independent admission review.

Dataset / taskMetricValueRevisionProvenance
agents-last-exam/agents-last-exam
overall-pass-rate
—20.3947
Unpinned
not published
Reported
Quarantined
agents-last-exam/agents-last-exam
overall-score
—42.8848
Unpinned
not published
Reported
Quarantined
agents-last-exam/agents-last-exam
ale-cli-pass-rate
—19.0476
Unpinned
not published
Reported
Quarantined
agents-last-exam/agents-last-exam
ale-cli-score
—43.5407
Unpinned
not published
Reported
Quarantined
agents-last-exam/agents-last-exam
near-term-pass-rate
—31.3433
Unpinned
not published
Reported
Quarantined
agents-last-exam/agents-last-exam
near-term-score
—60.0256
Unpinned
not published
Reported
Quarantined
agents-last-exam/agents-last-exam
full-spectrum-pass-rate
—21.8182
Unpinned
not published
Reported
Quarantined
agents-last-exam/agents-last-exam
full-spectrum-score
—42.5452
Unpinned
not published
Reported
Quarantined
agents-last-exam/agents-last-exam
last-exam-pass-rate
—2.6316
Unpinned
not published
Reported
Quarantined
agents-last-exam/agents-last-exam
last-exam-score
—14.7264
Unpinned
not published
Reported
Quarantined
allenai/IFBench_test
prompt_level_loose
—79.5
Unpinned
not published
Reported
Quarantined
cais/hle
hle
—30.8
Unpinned
not published
Reported
Quarantined
claw-eval/Claw-Eval
multimodal
—57.4
Unpinned
not published
Reported
Quarantined
datacurve/deep-swe
deep_swe
—42.2
Unpinned
not published
Reported
Quarantined
facebook/mmi
mmi-value
—11.46
Unpinned
not published
Reported
Quarantined
facebook/mmi
modality-presence-score
—18.09
Unpinned
not published
Reported
Quarantined
facebook/mmi
error-rate
—41.1
Unpinned
not published
Reported
Quarantined
facebook/mmi
pass-rate-strict
—9.18
Unpinned
not published
Reported
Quarantined
facebook/mmi
pass-rate-lenient
—13.33
Unpinned
not published
Reported
Quarantined
facebook/mmi
recall
—16.93
Unpinned
not published
Reported
Quarantined

Showing 20 of 143 rows.

Other versions (2)

ModelCap Index · Current board

Methodology
63.1
Measured on 4 public benchmarks
Current board
#68 of 371
Score range
58.5–67.7
Rank range
#51–#92
Evidence as of
18 Nov 2026, 14:36 UTC
Benchmarks used
ARC-AGI-2 Verified, Arena Coding, Arena Overall, LMArena Agent
Technical details
Public rank basis
Measured (measured)
Evidence label
Measured · 4 public benchmark observations across 4 boards
Index support
67.9% · strong
Identity binding
Exact catalogue product · qwen/qwen/qwen3.8-27b · aggregates disclosed benchmark configurations · endpoint configuration not claimed · artifact metadata Qwen/Qwen3.8-27B@37561816e66b5b9153b1b64c10968c18c1e6f4a0 (not the evaluation revision)
Index observations used
4
Qualified Index evidence point
63.1
Bayesian Evidence Score · measured evidence input only

BES normalizes admitted public benchmark evidence for measured Index rows. Its legacy score and observables prior do not define the public language rank.

Observed capability
63.1
Evidence status
confirmed · 90% mass
General preference
64.7
Coding
68.0
Agents & tools
32.5
Reasoning
38.1
Evidence breadth
100%
Evidence coverage
68%
Market and catalogue signals · secondary, never the language Index rank
Usage (OpenRouter popularity)
37.8
Liquidity (providers × uptime)
96.4
Open reach (HF downloads)
96.4
Surface (context / tools / modalities)
100.0
Freshness
69.2
  • · OpenRouter popularity #38
  • · 24 live providers
  • · HF 30d downloads 6888351
  • · Catalogue: context 1024000, reasoning, tools, modalities Text/Image/Video
  • · First seen 2026-08-14

These adoption and deployment observations remain context only. They do not change this model's ModelCap Index score or rank.

BES evidence selected · measured input only

The bounded 0–100 capability score blends each source’s competitive placement with its published achievement, then combines capability families. Evidence breadth adds a modest uncertainty adjustment; price and popularity are not part of benchmark-led rank.

Deployment readiness

Metadata index
93.2
strong · 94% metadata coverage
Not capability
Artifact reproducibility
85.0
Access & legal clarity
100.0
Deployability
99.3
Evaluation provenance
89.7

Not yet published: the base model it derives from, pinned evaluation dataset versions.

A metadata completeness and deployability index—not a safety certification, quality grade, or production-readiness claim.

Market activity

Venue coverage
Market Gravity
62.8
OpenRouter weekly popularity
#38 of 345

Market Gravity uses OpenRouter’s full-catalogue popularity order. Sparse Vercel top-ten observations appear only when published and do not affect the score.

Arena preference

Status
Ranked
Source
LMArena
Official rank
#112 of 422
Rating
1433.9
95% confidence interval
1429–1438
Votes
36K
Source published at
18 Nov 2026, 00:00 UTC
Category
overall
Identity match confidence
100%

Market Gravity breakdown

Usage55%
37.8

OpenRouter weekly popularity across the full model catalogue.

Liquidity25%
96.4

Independent providers versus the model's open or closed cohort, weighted by uptime.

Open reach15%
96.4

Hugging Face 30-day downloads for open models; neutral for closed models.

Freshness5%
69.2

Time since first public availability, on a six-month half-life.

Specification

Model ID
qwen/qwen3.8-27b
Context window
1M tokens
Declared output cap
236K tokens

OpenRouter catalogue or provider-endpoint declaration. Limits can vary by provider and prompt length; a model-wide maximum has not been independently verified. OpenRouter metadata

Inputs
Text, Image, Video
Outputs
Text
Tokenizer
Qwen
Cached input tokens
$0.059 / 1M
First seen on OpenRouter
14 Aug 2026

Capabilities

  • Supported: Reasoning
  • Supported: Tool use
  • Supported: Structured output
  • Supported: Response format
  • Not supported: Moderated

Weights & access

Open weights. Downloadable weights are published under an OSI or free-culture license in the current repository metadata. This describes weight availability, not whether the full training stack qualifies as Open Source AI.

Access
Open weights
Licence
apache-2.0
Pinned revision
37561816e66b
Parameters
27.8B
Architecture
Qwen3_5ForConditionalGeneration
Model type
qwen3_5
Hugging Face downloads (30d)
6.9M
Downloads (all time)
23.9M
Likes
21K
Repository updated
7 Nov 2026, 04:42 UTC

Compare Qwen3.8 27B with

Pick any model

Benchmark leaderboards featuring Qwen3.8 27B

All leaderboards
Public benchmark boards on which Qwen3.8 27B has a published result
LeaderboardScoreSource rankConfigurationPublished
Arena text leaderboardArena (LMArena)14341429–1438#112 of 422Not reported18 Nov 2026
Arena coding leaderboardArena (LMArena)14981491–1505#88 of 417Default18 Nov 2026
LMArena Agent leaderboardLMArena-5.5%-6.4%–-4.6%#51 of 68Default18 Nov 2026
ARC-AGI-2 leaderboardARC Prize Foundation42.4%#135 of 299X-High18 Nov 2026
Artificial Analysis Intelligence IndexArtificial Analysis20.1#83 of 125X-High28 Oct 2026
WildClawBench OpenClaw leaderboardWildClawBench48.0#15 of 34Default20 Jul 2026

Alternatives to Qwen3.8 27B, by the numbers

Ranked neighbours

Models within 3 places of #68 on the ModelCap Index.

Cheaper at a similar or higher Index

Index score within 3 points of, or above, this model's, with a lower listed blended price (3:1 input:output).

Other ranked Qwen models

The same publisher's other current models with a public Index position.

Qwen3.8 27B: common questions

How does Qwen3.8 27B rank among AI models?

Qwen3.8 27B holds ModelCap Index position #68 of 371 ranked language models as of 19 November 2026, with an Index score of 63.1 (interval 58.5–67.7); evidence: measured. The Index combines public benchmark boards with published uncertainty; the methodology page explains the weighting.

How much does Qwen3.8 27B cost per 1M tokens?

Qwen3.8 27B is listed at $0.06 per 1M input tokens and $2.20 per 1M output tokens as of 19 November 2026; the lowest output price among 24 providers is $1.05 on StreamLake. Prices come from the OpenRouter catalogue and re-observe on every refresh. For 1,000 requests using 2,000 input and 500 output tokens each, the listed rates imply $1.22 total (2M input + 0.5M output). This excludes caching, batch discounts, prompt-length tiers, tool charges and retries; verify the chosen endpoint's rate and limits before budgeting.

What is the context window of Qwen3.8 27B?

Qwen3.8 27B has a published context window of 1M tokens (1,024,000). The snapshot lists a 235,929 token catalogue-declared route output cap. OpenRouter catalogue or provider-endpoint declaration. Limits can vary by provider and prompt length; a model-wide maximum has not been independently verified. The providers table lists each endpoint's context limit.

Is Qwen3.8 27B open-weight?

Open weights (apache-2.0). Downloadable weights are published under an OSI or free-culture license in the current repository metadata. This describes weight availability, not whether the full training stack qualifies as Open Source AI. The repository is Qwen/Qwen3.8-27B on Hugging Face.

Which API providers serve Qwen3.8 27B?

24 providers serve Qwen3.8 27B through OpenRouter as of 19 November 2026: Darkbloom, Chutes, Near AI, Ionstream, Modal and DekaLLM and 18 more. Each provider's price, context limit, quantization and measured uptime are in the providers table above.

Which benchmarks has Qwen3.8 27B been evaluated on?

Qwen3.8 27B has published results on 6 public boards tracked by ModelCap as of 19 November 2026: Arena text 1434 (#112 of 422), Arena coding 1498 (#88 of 417), LMArena Agent -5.5% (#51 of 68), ARC-AGI-2 42.4% (#135 of 299) and AA Intelligence Index 20.1 (#83 of 125) and more. Each board page ranks every tracked model on that benchmark. The ModelCap Index scores the results its methodology admits, with published uncertainty, and shows reference boards without scoring them.

How does Qwen3.8 27B compare with Qwen3.5 397B A17B?

Qwen3.8 27B ranks #68 (Index 63.1) and Qwen3.5 397B A17B ranks #52 (Index 67.2) on the ModelCap Index as of 19 November 2026. The head-to-head page lines up their benchmarks, listed prices, context windows and provider counts side by side.

When was Qwen3.8 27B released?

Qwen3.8 27B first appeared in the catalogue on 14 Aug 2026. It is the current version in its family. Rank and price movements since then are recorded on the site's changes feed.