#DSpark
maybe.

ooooo also, there’s a DSpark SD module so it’s fucking fast too
November 20, 2026 at 9:49 PM
效率数字也值得看:模型算力利用率 37%,而同类开源训练器只有 16%,同样训练量省 57% 的 GPU 时间;端到端快了 2 倍。

怎么用:权重 Apache 2.0 协议,BF16、FP8、NVFP4 三种精度都在 Hugging Face 上,保留了 Qwen3.8-27B 的架构和 25.6 万 token 上下文。官方推荐用 SGLang + DSpark 推测解码来跑,模型默认先思考再回答,推理过程放在 reasoning_content 字段里。
November 19, 2026 at 7:42 AM
what you're describing is speculative decoding of the DSpark/DFlash variety

Sparse Attention is the technique of being selective about what tokens can attend to what others (normally, every token attends to every one behind it, at least in decoder-only models (which is ~everything these days))
November 16, 2026 at 6:37 AM
Liquid AI accelerates vision-language models with LFM2.5-VL-DSpark. The team achieves major speed gains through distributed inference and optimized attention mechanisms. This makes multimodal AI more practical for real world deployment at scale.
Accelerating vision-language models with LFM2.5-VL-DSpark
A Blog post by Liquid AI on Hugging Face
huggingface.co
November 12, 2026 at 9:49 PM
Offline-only limits fine-grained latency tuning for DSpark. Bare-metal deployment needs a persistent container or systemd service instead of ephemeral state to prevent context leaks and ensure session stability under heavy concurrency.
November 10, 2026 at 2:22 AM
4× DGX Spark (GB10s capped at 2,200 MHz): DeepSeek V4.1 Flash 1017 at 116.0 tok/s writing code (middle of 3, 112.6–120.2) and 4,924 tok/s reading ~15k tokens. DSpark on.
getrigbase.com/builds/4x-dg...
November 8, 2026 at 7:23 AM
DeepSeek open-sourced DeepInfer for V4.1-Flash, but it's offline-only.

Understand DSpark decoding, why vLLM isn't a replacement, and how to deploy on bare-metal SM100 GPUs:

www.servermo.com/howto/deploy...

#DeepSeek #AI #MLOps #DevOps #SRE
November 6, 2026 at 7:23 AM
TensorRT Edge-LLM으로 DGX Spark에서 Nemotron 3.5 Lightning 실행하기 (DSpark로 생성 속도 최대 1.7배)

핵심 요약

• NVIDIA가 기술 블로그 튜토리얼에서 DGX Spark 64GB에 TensorRT Edge-LLM 0.11.1을 설치해 Nemotron 3.5 Lightning(전체 30B, 활성 3B, NVFP4)을 실행하는 절차를 공개했습니다.
• trtedge chat 명령 하나로 약 21GB 모델 다운로드, 엔진 빌드, 채팅이 이어지고, trtedge serve로 Open…
TensorRT Edge-LLM으로 DGX Spark에서 Nemotron 3.5 Lightning 실행하기 (DSpark로 생성 속도 최대 1.7배)
핵심 요약 • NVIDIA가 기술 블로그 튜토리얼에서 DGX Spark 64GB에 TensorRT Edge-LLM 0.11.1을 설치해 Nemotron 3.5 Lightning(전체 30B, 활성 3B, NVFP4)을 실행하는 절차를 공개했습니다. • trtedge chat 명령 하나로 약 21GB 모델 다운로드, 엔진 빌드, 채팅이 이어지고, trtedge serve로 OpenAI 호환 API 서버를 띄웁니다(문서상 실험 단계, 기본 동시 요청 1개). • NVIDIA 측정(배치 1, 출력 128토큰 제한)에서 DSpark 추측 디코딩은 생성 속도를 1.45~1.7...
discuss.pytorch.kr
November 4, 2026 at 10:35 AM
DeepSeek just open-sourced the inference engine behind its production services, with optimized kernels and DSpark draft-token decoding. 🔗 github.com/deepseek-ai/DeepInfer

Credits: github.com/deepseek-ai/DeepInfer

#deepinfer #deepseek #python
November 2, 2026 at 11:30 PM
Liang Wenfeng's name on an open-weights release is the strongest signal in Chinese AI this month. Still, a ten-point explainer is not an eval, and nothing here is independently benchmarked. Hard to see it unseating Qwen on local rigs until someone publishes numbers.
DSpark, with Liang Wenfeng’s name on it: ten points are all you need to understand it
A ten-point explainer on DeepSeek’s DSpark speculative-decoding system, arguing that its real breakthrough is systems engineering: combining parallel draft generation with lightweight serial correction and hardware-aware scheduling to deliver 60–85% single-user speed-ups and up to 4× effective throughput under high concurrency.
aichina.news
November 3, 2026 at 8:08 PM
769 tok/s on RTX 4090! Spark-X2.6-1.7B DSpark (GSM8K, Q4_K_M, llama.cpp).
Up to 2.43x speedup (BF16, GSM8K).
Like & download:
https://modelscope.cn/models/XHToken/Spark-X2.6-1.7B-DSpark-GGUF
https://huggingface.co/XHToken/Spark-X2.6-1.7B-DSpark-GGUF
November 1, 2026 at 1:39 PM
DSpark for Spark-X2.6-4B is here! ✨
Average BF16 speedup: 2.14x RTX 4090 / 2.50x M5 Pro.
❤️ Like & download!
ModelScope: https://modelscope.cn/models/XHToken/Spark-X2.6-4B-DSpark-GGUF
https://huggingface.co/XHToken/Spark-X2.6-4B-DSpark-GGUF
October 31, 2026 at 6:58 AM
NVIDIA: TENSORRT EDGE-LLM RUNS NEMOTRON-3.5 LIGHTNING ON DGX SPARK 64GB, 126.9 TOKENS/SEC WITH DSPARK
October 31, 2026 at 3:00 PM
NVIDIA: TENSORRT EDGE-LLM RUNS NEMOTRON-3.5 LIGHTNING ON DGX SPARK 64GB, 126.9 TOKENS/SEC WITH DSPARK
October 31, 2026 at 3:00 PM
Eight H200s ran FP8 in vLLM with DSpark, which drafts tokens for the model to check. Generation was about 65 tok/s for one 1K-token prompt, 120 to 138 on chat with accepted drafts, and about 480 total across 1,000 conversations. The recipe allows up to 16 requests at once.
October 31, 2026 at 5:21 PM
Big speedup multipliers flatter slow setups.

SparkLLM's new DSpark draft model, announced on X, advertises 2.50x on an M5 Pro. That's the full-precision build, ending at 77 tok/s. The 4-bit build gains much less and still lands at 115. I'd pick on final speed.
October 31, 2026 at 1:15 PM
the second gpu can hold things like the dspark drafter and vision encoder though
October 29, 2026 at 5:49 PM
Strata 0.2.7, released 27 October, shipped a Windows engine for Strix Halo's integrated Radeon GPU. The developer used a GMKtec EVO-X2 with a Ryzen AI Max+ 395, the ISTA-DASLab IQ2_XXS set at 2.00 bits per weight and the DSpark drafter.
October 30, 2026 at 6:08 PM
MiniCPM's dense-document OCR at small size was promising, so a dedicated OCR-1.0 makes sense. But no model card, no parameter count, no benchmark — hard to recommend this over PaddleOCR in production until weights and evals land.
OpenBMB Lists MiniCPM-OCR-1.0 and DSpark Variant on Model Hubs
OpenBMB has published model pages for MiniCPM-OCR-1.0 and MiniCPM-OCR-1.0-DSpark on Hugging Face, ModelScope and Modelers. The Hugging Face listing for the DSpark variant gives its task as image-text-to-text.
aichina.news
October 29, 2026 at 2:34 PM
⚡ Embedded DSpark speculative decoding head: up to 1.9x throughput, same outputs. NVFP4 checkpoint for Blackwell GPUs. Serve with vLLM, SGLang, Transformers or llama.cpp; fine-tune with Axolotl, Unsloth or TRL
October 29, 2026 at 5:45 PM
MiniCPM OCR family 🔥 1st OCR from OpenBMB

✨ MiniCPM OCR 1.0
- 1.3B
- Apache 2.0
- 55% higher processing efficiency with DSpark
- Handles complex layouts, tables & formulas
October 27, 2026 at 6:06 PM
OPENBMB: MINICPM-OCR-1.0-DSPARK ON HUGGING FACE, 55% FASTER DECODING, 296M PARAMS
October 27, 2026 at 6:15 PM
October 27, 2026 at 6:15 PM
I realized that DeepSeek models' DSpark is trained on DeepSeek Harness. So the output token can be 500 tps instead of 50 tps like I'm using a Pi agent. Which is a torture to work with, takes forever, and messing up my work. DeepSeek Harness is actually cautious about scope creep; it'd stop and ask.
To be real, I haven't found an autonomous software that can limit the context window to 75k tokens like Pi. And switching out faster models to compact the context. So I'm stuck in Pi, which shoves everything into the screen instead of delegating.
See? The chatbot on benchmark memorization has no common sense at all. It couldn’t handle a task that is so simple. Guess what? It picked GPT 6.1 Sol as extractive QA for web scraping. That's insane. The owner has 6 months to figure this out...
October 27, 2026 at 3:29 AM