Skip to content
Tier A — Frontier
Runs in:FranceMade in:China
OVH AI Endpoints (GRA)

Qwen3.5-397B-A17B

Tier A — Frontier

Tokonomix Editorial Team·Reviewed by Mes Kalkan·
Section 01

Speed analysis

Latency measured across all benchmark runs. P50 (median) and P95 (95th percentile) give a realistic picture of response speed under normal and peak load.

P50 latency (median)P95 latency100 runs
149256049717382979308-2109-14ms
Section 02

Quality scores

How this model compares to the rest of the field on each prompt category, from a pairwise fit over the same prompts. The raw judge score sits underneath each number.

42%
Coding
judge mean 89
6%
Creative
judge mean 43
6%
Factual
judge mean 14
28%
Multilingual
judge mean 38
5%
Reasoning
judge mean 1

Win rate per category: how often this model beats a field-average model on a prompt from that category. 50% is average, not a failing grade. It is not a percentage of correct answers.

Section 03

Pricing

What you pay per million tokens when you use this model on Tokonomix, plus an estimate for a typical conversation.

💰
API rates — Qwen3.5-397B-A17B
$1.07 per 1M input tokens
$6.36 per 1M output tokens
≈ $0.0019 per typical conversation (800 tokens)
Input vs output price (per 1M tokens)
per 1M input tokens$1.07
per 1M output tokens$6.36
Section 04

Tokens per second

Throughput in tokens per second, derived from measured P50 latency. Higher is better; fluctuations track provider-side load.

Throughput (tokens / s)625 / avg 865
132651

Estimated from P50 latency × 200 output tokens — the absolute number depends on this assumption; the trend is what matters.

Section 05

Capabilities

ownedBy: Qwen
Section 06

Availability

Availability

How often this model answers when we call it — measured across real API requests and live tests over the last 30 days. This is separate from quality: these numbers only tell you whether the model responds, not how good the answer is.

Last 7 days

100.0%

n=16

Last 30 days

100.0%

n=16

Median response time

17,684ms

n=16

Based on 401 measurements over the last 30 days.

Technical details

Only live API calls and live-test requests count — internal probes and benchmark runs are excluded.

Calls with a custom API key (BYOK) are excluded: those failures are key-specific, not a sign of model downtime.

Failed calls are NOT included in quality scores — quality is measured on successful responses only. Availability and quality are independent signals.

Median response time (p50) across successful calls with a recorded duration. Outliers (very slow or very fast calls) pull the median less than the average.

Total calls (30d)

16

OK responses (30d)

16

Total calls (7d)

16

OK responses (7d)

16

Section 07

Tokonomix benchmark verdicts

⚖️
Endorsed by 2 judges
Independent LLM judges evaluated this model on our weekly intelligence tests
cohere/command-a100/100 · 1 runs
1 correct0 partial0 wrong100% accuracy
claude-sonnet-4-543/100 · 69 runs
24 correct3 partial42 wrong35% accuracy
2026-09-13

Qwen3.5-397B-A17B quality drops 25 points with slower response times

Qwen3.5-397B-A17B has experienced a notable performance decline in this benchmark window. The overall quality score dropped from a perfect 100 to 75 points, a 25-point decrease that represents a significant regression. The coding category score mirrors this decline, falling from 100 to 75. Response times have also degraded, with the median latency increasing by 34 percent from 3362ms to 4505ms. This means users can expect responses to take approximately 1.1 seconds longer on average. The combination of quality degradation and slower performance suggests potential infrastructure issues or model serving problems that warrant investigation. With only two test runs in the current window compared to one in the previous period, the sample size remains limited but the trend is concerning. Users relying on this endpoint for coding tasks should be aware of the reduced quality scores and prepare for longer wait times. The model appears to have shifted from excellent performance to merely adequate results across measured dimensions.

Quality

75.0

Latency p50

4,505 ms

Test runs

2

Quality dropped 25 points Latency increased 34% Coding score fell to 75 Response times now 4.5 seconds
Section 08

Full model profile

Qwen3.5-397B-A17B: a 397-billion-parameter model that only wakes up 17 billion at a time

Qwen3.5-397B-A17B is Alibaba's Qwen team's follow-up to the Qwen3 generation, released as open weights under an Apache 2.0 license in early 2026. The name states the architecture directly: 397 billion total parameters, with roughly 17 billion active for any given token. It is a sparse Mixture-of-Experts model, and Qwen positions it as a unified vision-language foundation — a single model trained from the start on text, images, and video together, rather than a text model with a vision adapter bolted on afterward.

Under the hood

According to Qwen's own model documentation, the architecture pairs Gated DeltaNet — a linear-attention mechanism — with sparse MoE routing across 60 layers, arranged as repeating blocks of linear-attention-plus-MoE with periodic full-attention layers mixed in. The MoE layer itself carries 512 experts, of which 11 are activated per token: 10 routed experts chosen dynamically plus one shared expert that is always active. That combination — mostly-linear attention for efficiency, MoE for capacity, occasional full attention for the kind of long-range dependency linear attention alone struggles with — is what lets a model with 397 billion total parameters run at roughly the compute cost of a 17-billion-parameter dense model per token.

Context length is 262,144 tokens natively, and Qwen documents extension up to roughly 1,010,000 tokens via YaRN scaling — a technique that adjusts the model's positional encoding to generalize beyond its trained context length, typically with some quality trade-off the further past the native window a request pushes.

Where it shines

Qwen reports strong results on its own benchmark suite: 87.8 on MMLU-Pro, 70.4 on SuperGPQA, and 94.8 on the HMMT February 2025 math competition set, alongside 85.0 on MMMU and 86.7 on MLVU for image and video understanding respectively. These are Qwen's own published figures rather than independently reproduced numbers, so they should be read as the vendor's account of the model's own progress — but taken together they describe a model built for broad, general-purpose reasoning and multimodal understanding rather than a narrow specialist.

The open-weights license is itself a practical advantage: Apache 2.0 permits self-hosting, fine-tuning, and redistribution without the licensing friction that comes with a closed API-only model, which matters for teams that need to run inference on their own infrastructure or adapt the model to a narrow domain.

Where it falls short

Qwen's own documentation frames this release as "cross-generational parity with Qwen3" on many dimensions — meaning the gains over the prior generation are presented as broad and incremental rather than a step-change, and language in the model card is careful not to claim a dramatic leap. As with any Mixture-of-Experts model, the discrepancy between total and active parameters means the model needs far more memory to hold in full than its per-token compute cost would suggest — 397 billion parameters have to be resident somewhere even though only a fraction fire on any given request, which shapes what hardware self-hosting actually requires.

When to pick it

This model fits general-purpose reasoning, coding, and multimodal workloads where a team wants either the option to self-host or the cost profile of a very large model without paying full dense-model compute per request. It is a reasonable default for teams already invested in the Qwen ecosystem, or for workloads that need a genuinely large native context window without leaning on aggressive scaling tricks.

Alternatives worth comparing

The predecessor Qwen3 generation remains available for teams that do not need the added vision and video understanding this release adds. For workloads that need a smaller footprint, Qwen has also shipped considerably smaller models in the same family line that trade some capability for a much lighter memory footprint — worth benchmarking directly against this model on the specific task before committing to the larger one.

Last automated test
Sep 14, 2026 · 20:02 UTC · Speed benchmark
P50 latency
320 ms
P95 latency
400 ms
Errors
0 / 6 runs
Last reviewed by Tokonomix Team·September 14, 2026