Primary use stated or inferred cautiously from official documentation.
SingProbe Qwen3.8-Flash-Next
Trained inclusionAI probe that reuses its base model's hidden states to classify intent, response unsafety and hallucination risk during generation. Its own checkpoint was published on October 27, 2026. It is not a standalone assistant or a new base-model version. Its ten outputs cover eight intent categories and two risk signals.
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
Limits and integration
| API ID | inclusionAI/Qwen3.8-Flash-Next-singprobe |
|---|---|
| Model type | Text classification |
| Access model | Open weights |
| License | Qwen Community License 1.0 |
| Deployment | Base model and adapted runtime |
| Release | 27 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Hidden states |
| Exit | Per-token scores |
| Context window | Depends on the base model and runtime |
| maximum output | Ten scores per token |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · SGLang · vLLM |
Monitoring per-token risks in systems already serving the compatible base model. The probe weights are small relative to the model being observed, but do not replace it. The client receives label_0 through label_9 using the integrations listed in the card; it must choose thresholds and how to act on these signals.
Requires the exact base-model/probe pair and a modified runtime. It cannot accept a conversation on its own and does not guarantee the absence of hallucinations. Overhead below 0.5% is reported by the provider, not promised for all hardware. Do not confuse its scores with a general language benchmark.
What it can do
Base model
Qwen/Qwen3.8-Flash-Next
Probe parameters
5,18 M
False positives on benign responses
0,03 %
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No API price is published for the probe. Costs include the base model, memory, runtime and monitoring. The card refers to the base model's license; the Qwen variant retains Qwen Community License 1.0, with specific conditions for commercial services. The GLM variant links to MIT. Public downloads do not remove these obligations.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
How to read the results
The card reports intent F1 across six benchmarks, safety F1 across eight, streaming safety R-AUC and T-AUC across three, and hallucination AUC across six. These average different suites; they should not be added together or compared with coding accuracy. False positives are averaged across five datasets. Not reproduced by Inferama.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| SingProbe Query Intent | 0,8728 | F1 | View source ↗ |
| SingProbe Response Safety | 0,8714 | F1 | View source ↗ |
| SingProbe Streaming R-AUC | 0,9882 | R-AUC | View source ↗ |
| SingProbe Streaming T-AUC | 0,9320 | T-AUC | View source ↗ |
| SingProbe Hallucination | 0,8063 | AUC | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
Chronology
SingProbe Qwen3.8-Flash-Next
Version added to Inferama's verified catalog.