Alibaba
Qwen 3.8 27B
Qwen 3.8 27B is a model with open weights from Alibaba. On the Vals AI Legal Research Bench it reaches 36.06 percent all-pass, position 48 of 91, within a standard error of ±3.34 percentage points at $2.16 per task, as at 9 November 2026 [1]. It does not pass the compliance gate: Open weights, but no route found at Vertex AI (EU), Azure Foundry (Data Zone EU), AWS Bedrock (EU regions) or the Mistral EU endpoint: the gate only opens on a dedicated EU endpoint or self-hosting, where the region is fixed by contract (as in the existing open-weights assessment).
- Status unknown
- Open weights
- 14 August 2026
- PrudentScore
- -No score: the gate is not open
- Compliance gate
- RejectedOpen weights, but no route found at Vertex AI (EU), Azure Foundry (Data Zone EU), AWS Bedrock (EU regions) or the Mistral EU endpoint: the gate only opens on a dedicated EU endpoint or self-hosting, where the region is fixed by contract (as in the existing open-weights assessment)
- Cost per task
- $2.16Vals AI, per completed task [1]
01
Compliance gate
Why
- Open weights, but no route found at Vertex AI (EU), Azure Foundry (Data Zone EU), AWS Bedrock (EU regions) or the Mistral EU endpoint: the gate only opens on a dedicated EU endpoint or self-hosting, where the region is fixed by contract (as in the existing open-weights assessment).
The gate only opens when all three requirements are covered: processing in the EU, a contracting party in the EU or EEA, and a contractual exclusion of training on customer data. The gate tests a route, not a model name: the same model can pass through one provider and fail through another. Where a provider publishes no figure, we give no estimate.
Sources for this verdict
- docs.cloud.google.com/gemini-enterprise-agent-platform/resources/data-residency
- learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability
- docs.aws.amazon.com/bedrock/latest/userguide/models-region-compatibility.html
- docs.mistral.ai/inference/regional-inference
02
Routes and EU processing
The gate tests a route (processor, region and contract), not a model name. Per route: processing in the EU, the contracting party and training on customer data.
| Processor | Region or deployment | EU processing | Source |
|---|---|---|---|
| Alibaba Cloud Model Studio (Duitsland/Frankfurt, EU-bereik) | region Germany (Frankfurt) with service deployment scope EU ("inference within the region"); availability of this model in that region has not been verified | Unknown | SourceRead 10 November 2026 |
03
Scores per benchmark
Every measurement with its source, measurement date and the day we read it. The position counts only the models measured on the same benchmark and version.
- Legal Research BenchVersion 1 · Counts towards the PrudentScore36.06%± 3.3448 of 91
Vals AI, Legal Research Bench, v1, 9 November 2026 · Source · Measured 9 November 2026 · Read 10 November 2026 · cost per task $2.16
- Harvey LAB (Vals implementation)Version 1 · Counts towards the PrudentScore11.25%± 2.4215 of 80
Vals AI, Harvey LAB (Vals implementation), v1, 18 October 2026 · Source · Measured 18 October 2026 · Read 10 November 2026 · cost per task $4.03
- LegalBench (Vals leaderboard)Version 182.43%± 0.4767 of 113
Vals AI, LegalBench (Vals leaderboard), v1, 13 October 2026 · Source · Measured 13 October 2026 · Read 10 November 2026
- CaseLaw v2Not measured
- TaxEval v2Version 270.85%± 0.8977 of 108
Vals AI, TaxEval v2, v2, 1 September 2026 · Source · Measured 1 September 2026 · Read 10 November 2026
- Tax Agent BenchVersion 130.89%± 3.3231 of 82
Vals AI, Tax Agent Bench, v1, 9 November 2026 · Source · Measured 9 November 2026 · Read 10 November 2026 · cost per task $0.82
- CorpFin v2Not measured
- Immigration BenchNot measured
- Access to JusticeNot measured
- Public Benefits BenchNot measured
- Vals Legal IndexNot measured
- Vals IndexNot measured
- GPQA Diamond (Vals-run)Version 188.89%± 1.5832 of 109
Vals AI, GPQA Diamond (Vals-run), v1, 1 September 2026 · Source · Measured 1 September 2026 · Read 10 November 2026
- MMLU Pro (Vals-run)Version 184.34%± 0.3665 of 110
Vals AI, MMLU Pro (Vals-run), v1, 1 September 2026 · Source · Measured 1 September 2026 · Read 10 November 2026
- Artificial Analysis Intelligence IndexNot measured
Source: Vals AI (Legal Research Bench, Harvey LAB (Vals implementation), LegalBench (Vals leaderboard), TaxEval v2, …) · Licence: CC BY 4.0 · Adaptation: Prudai selected the models, translated the explanations, added the gate verdict and the PrudentScore and rounded figures to two decimals.
The PrudentScore is a composition by Prudai B.V. based on publicly published figures and does not represent the view or endorsement of Vals AI, Harvey or Artificial Analysis. Figures from Artificial Analysis appear in running text only, with attribution and a link, as that party's terms require [9].
04
Key facts
- Developer
- Alibaba
- Model family
- Qwen
- Release date
- 14 August 2026
- Status
- Status unknown
- Context window
- 262K tokens
- Max output tokens
- 66K tokens
- List price
- $0.50 input · $3.00 outputper million tokens · Read 10 November 2026
- Weights
- Open
- = not yet recorded in the register.
05
Frequently asked questions about this model
Does Qwen 3.8 27B pass the compliance gate for legal work?
No. Open weights, but no route found at Vertex AI (EU), Azure Foundry (Data Zone EU), AWS Bedrock (EU regions) or the Mistral EU endpoint: the gate only opens on a dedicated EU endpoint or self-hosting, where the region is fixed by contract (as in the existing open-weights assessment). The gate tests a route, not a model name; the sources are on this page.
What does Qwen 3.8 27B cost per task?
Vals computes $2.16 per completed task on the Legal Research Bench [1]. The list price is $0.50 per million input tokens and $3.00 per million output tokens.
How does Qwen 3.8 27B score on legal research?
36.06 percent all-pass on the Vals AI Legal Research Bench, position 48 of 91 measured models, as at 9 November 2026 [1]. That benchmark measures US law; an independent, published benchmark for Dutch legal work does not exist yet.
Related models
Claude and Anthropic, GPT and OpenAI, Gemini and Google, Grok and xAI, Kimi and Moonshot AI, Muse and Meta, Mistral, DeepSeek, GLM and Z.ai, Harvey, Vals AI and Artificial Analysis are trademarks of their respective owners. They are named for identification only and this implies no partnership, sponsorship or endorsement.
Brand marks appear for identification only and imply no partnership, sponsorship or endorsement.