Primary use stated or inferred cautiously from official documentation.
T
TENCENT / tencent/Qwen3.8-27B-DSpine-b16
DSpine for Qwen3.8-27B
Tencent component that drafts tokens in parallel to accelerate Qwen3.8-27B inference. It has five layers and about 2.026 billion BF16 parameters; 27B names the target model. Weights were published on November 7. It is not a standalone assistant: the target model verifies candidates before accepting them into the output.
01 / SUMMARY
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.
02 / TECHNICAL SHEET
Limits and integration
| API ID | tencent/Qwen3.8-27B-DSpine-b16 |
|---|---|
| Model type | Inference acceleration |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted |
| Release | 07 NOV 2026 |
| Knowledge cutoff | Not published |
| Entrance | Target-model tokens and states |
| Exit | Candidate tokens |
| Context window | 8,192-token sliding window in the draft |
| maximum output | Blocks of 16 candidate tokens; target-model verification |
| Reasoning | Depends on the target model |
| Published tools | SGLang 0.5.22 · FlashAttention 3 · ReplaySSM |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · SGLang 0.5.22 |
Integrating speculative decoding into a self-hosted target-model deployment. The official code uses SGLang 0.5.22, FlashAttention 3, dynamic verification and ReplaySSM. The draft window is not the service’s total context. The example configures 131,072 tokens for the combined system, with up to eight concurrent requests; this is not a universal limit for every installation.
Results depend on the target model, runtime, precision and workload. Measurements replaying tool-decision prefixes do not show that an agent completes real tasks better. The draft is not credited with conversational capabilities, vision, autonomous tool use or improved accuracy. The card does not identify the GPU used for its tables, so it cannot promise the same speedup on other hardware.
03 / CAPABILITIES
What it can do
Orientation
Integrating speculative decoding into a self-hosted target-model deployment. The official code uses SGLang 0.5.22, FlashAttention 3, dynamic verification and ReplaySSM. The draft window is not the service’s total context. The example configures 131,072 tokens for the combined system, with up to eight concurrent requests; this is not a universal limit for every installation.
Context
8,192-token sliding window in the draft · Blocks of 16 candidate tokens; target-model verification
Tools and integration
SGLang 0.5.22 · FlashAttention 3 · ReplaySSM
Access
Hugging Face · requires Qwen3.8-27B
Input modalities
Target-model tokens and states
04 / PRICES
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 |
The release provides Apache 2.0 weights, not a commercial per-token API. Running them also requires the target model and compute resources. Unknown rates remain unpublished; they are not presented as free, and throughput is not used to calculate savings.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Tencent measures acceptance length and system throughput, not answer accuracy. It reports 7.20 accepted tokens per iteration on GSM8K at temperature zero. On SWE-bench Pro, with low reasoning and concurrency eight, it reports 713.41 tokens/s versus 241.86 without speculation, a stated ratio of 2.95 times. The test replays tool-decision prefixes and allows up to 32,768 new tokens per request. Inferama has not reproduced these measurements.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| DSpine Qwen3.8-27B · GSM8K acceptance | 7,20 | Accepted tokens per iteration | View source ↗ |
| DSpine Qwen3.8-27B · SWE-bench Pro throughput | 713,41 | Output tokens per second | 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.
06 / VERSIONS
Chronology
Qwen3.8-27B-DSpine-b16
Auxiliary BF16 checkpoint released; requires the target model.
07 / SOURCES
Traceability
TECHNICAL SHEETOfficial checkpoint card · pinned revisionhuggingface.co↗
IMPLEMENTATIONOfficial code and target-model integrationgithub.com↗
CONFIGURATIONOfficial draft-model configurationhuggingface.co↗
DISPONIBILIDADWeights and files of the verified revisionhuggingface.co↗
LICENSEProject Apache 2.0 licensegithub.com↗