軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
https://gigazine.net/news/20260925-lfm2-5-vl-dspark/
https://gigazine.net/news/20260925-lfm2-5-vl-dspark/
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
アメリカに拠点を置くAI開発企業のLiquid AIが視覚言語モデル「LFM2.5-VL-3B-DSpark」を2026年9月24日に公開しました。LFM2.5-VL-3B-DSparkは「LFM2.5-VL-3B」に投機的デコード技術「DSpark」を適用したモデルで、性能を維持したまま数倍の高速化を実現しています。
gigazine.net
September 25, 2026 at 4:50 AM
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
https://gigazine.net/news/20260925-lfm2-5-vl-dspark/
https://gigazine.net/news/20260925-lfm2-5-vl-dspark/
Source, this entire video is great
youtu.be/fN43F9n1mBs?...
youtu.be/fN43F9n1mBs?...
(YTP) Wario Kart World - Direct
YouTube video by Dspark
youtu.be
May 24, 2025 at 12:35 AM
Source, this entire video is great
youtu.be/fN43F9n1mBs?...
youtu.be/fN43F9n1mBs?...
DeepSeek-V4-Flash can now run 2× faster locally with DSpark! ⚡️
DSpark enables V4-Flash-0731 GGUFs to generate ~1.4–2× faster with no accuracy change.
DeepSeek-V4-Flash-0731 can reach at 120 tokens/s.
GGUFs: huggingface.co/unsloth/Deep...
Guide: unsloth.ai/docs/models/...
DSpark enables V4-Flash-0731 GGUFs to generate ~1.4–2× faster with no accuracy change.
DeepSeek-V4-Flash-0731 can reach at 120 tokens/s.
GGUFs: huggingface.co/unsloth/Deep...
Guide: unsloth.ai/docs/models/...
August 6, 2026 at 2:14 PM
DeepSeek-V4-Flash can now run 2× faster locally with DSpark! ⚡️
DSpark enables V4-Flash-0731 GGUFs to generate ~1.4–2× faster with no accuracy change.
DeepSeek-V4-Flash-0731 can reach at 120 tokens/s.
GGUFs: huggingface.co/unsloth/Deep...
Guide: unsloth.ai/docs/models/...
DSpark enables V4-Flash-0731 GGUFs to generate ~1.4–2× faster with no accuracy change.
DeepSeek-V4-Flash-0731 can reach at 120 tokens/s.
GGUFs: huggingface.co/unsloth/Deep...
Guide: unsloth.ai/docs/models/...
DeepSeek DSpark for DeepSeek-V4 Flash and Pro - 60% to 85% faster generation?
Models: huggingface.co/deepseek-ai/...
Paper: github.com/deepseek-ai/...
Repo: github.com/deepseek-ai/...
Models: huggingface.co/deepseek-ai/...
Paper: github.com/deepseek-ai/...
Repo: github.com/deepseek-ai/...
June 27, 2026 at 5:04 PM
DeepSeek DSpark for DeepSeek-V4 Flash and Pro - 60% to 85% faster generation?
Models: huggingface.co/deepseek-ai/...
Paper: github.com/deepseek-ai/...
Repo: github.com/deepseek-ai/...
Models: huggingface.co/deepseek-ai/...
Paper: github.com/deepseek-ai/...
Repo: github.com/deepseek-ai/...
THAT'S WHAT I'M SAYING!!! Oh also see can see now and with N-gram + dspark it should FUCKING RIP!!!! Like, I am really interested in seeing the TPS...
HOLY FUCK!!!
HOLY FUCK!!!
September 10, 2026 at 6:36 AM
THAT'S WHAT I'M SAYING!!! Oh also see can see now and with N-gram + dspark it should FUCKING RIP!!!! Like, I am really interested in seeing the TPS...
HOLY FUCK!!!
HOLY FUCK!!!
Run Gemma-4-12B-it-8bit x1.5-2.5 faster on #Apple M4 Pro with DSpark/DFlash.
#LLM #AI
#OpenSource MIT lic
github.com/ARahim3/mlx-...
#LLM #AI
#OpenSource MIT lic
github.com/ARahim3/mlx-...
GitHub - ARahim3/mlx-dspark: DeepSeek's DSpark and z-lab's DFlash speculative decoding — running natively on Apple Silicon via MLX.
DeepSeek's DSpark and z-lab's DFlash speculative decoding — running natively on Apple Silicon via MLX. - ARahim3/mlx-dspark
github.com
July 11, 2026 at 8:25 PM
Run Gemma-4-12B-it-8bit x1.5-2.5 faster on #Apple M4 Pro with DSpark/DFlash.
#LLM #AI
#OpenSource MIT lic
github.com/ARahim3/mlx-...
#LLM #AI
#OpenSource MIT lic
github.com/ARahim3/mlx-...
GIGAZINE より
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
"LFM2.5-VL-3B-DSpark" exploded by applying DSpark's draft model to lightweight visual language model "LFM2.5-VL-3B"
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
"LFM2.5-VL-3B-DSpark" exploded by applying DSpark's draft model to lightweight visual language model "LFM2.5-VL-3B"
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場 - GIGAZINE / LFM2.5-VL-3B-DSpark - GIGAZINE Lightweight Visual Language Model LFM2.5-VL-3B-DSpark - GIGAZINE
gigazine.net
September 25, 2026 at 5:32 AM
GIGAZINE より
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
"LFM2.5-VL-3B-DSpark" exploded by applying DSpark's draft model to lightweight visual language model "LFM2.5-VL-3B"
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
"LFM2.5-VL-3B-DSpark" exploded by applying DSpark's draft model to lightweight visual language model "LFM2.5-VL-3B"
LiquidAI Releases LFM2.5-VL-DSpark for Vision-Language Models
https://localmodelwatch.tsuchitsuchi.com/en/2026/09/25/liquidai-lfm2-5-vl-dspark-released/
https://localmodelwatch.tsuchitsuchi.com/en/2026/09/25/liquidai-lfm2-5-vl-dspark-released/
September 25, 2026 at 7:09 AM
LiquidAI Releases LFM2.5-VL-DSpark for Vision-Language Models
https://localmodelwatch.tsuchitsuchi.com/en/2026/09/25/liquidai-lfm2-5-vl-dspark-released/
https://localmodelwatch.tsuchitsuchi.com/en/2026/09/25/liquidai-lfm2-5-vl-dspark-released/
ビジョン言語モデル向け高速化ドラフトモデル「LFM2.5-VL-DSpark」公開
https://localmodelwatch.tsuchitsuchi.com/2026/09/25/lfm2-5-vl-dspark/
https://localmodelwatch.tsuchitsuchi.com/2026/09/25/lfm2-5-vl-dspark/
September 25, 2026 at 7:09 AM
ビジョン言語モデル向け高速化ドラフトモデル「LFM2.5-VL-DSpark」公開
https://localmodelwatch.tsuchitsuchi.com/2026/09/25/lfm2-5-vl-dspark/
https://localmodelwatch.tsuchitsuchi.com/2026/09/25/lfm2-5-vl-dspark/
✓ 軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
- https://gigazine.net/news/20260925-lfm2-5-vl-dspark/
- https://gigazine.net/news/20260925-lfm2-5-vl-dspark/
September 25, 2026 at 5:03 AM
✓ 軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
- https://gigazine.net/news/20260925-lfm2-5-vl-dspark/
- https://gigazine.net/news/20260925-lfm2-5-vl-dspark/
DwarfStar in the latest two weeks was improved in almost every aspect for Metal, DGX Spark and Strix Halo. It is simpler to say: update, you will hopefully see speed and correctness improvements in many areas. Also DSpark with DeepSeek v4 Flash now works much better overall.
September 6, 2026 at 10:17 PM
DwarfStar in the latest two weeks was improved in almost every aspect for Metal, DGX Spark and Strix Halo. It is simpler to say: update, you will hopefully see speed and correctness improvements in many areas. Also DSpark with DeepSeek v4 Flash now works much better overall.
Accelerating Vision-Language Models with LFM2.5-VL-DSpark
What is LFM2.5-VL-DSpark? LFM2.5-VL-DSpark is the latest generation of vision-language models that combines visual perception and textual understa
https://bloggersminds.com/post/accelerating-vision-language-models-with-lfm2-5-vl-dspark-6547
What is LFM2.5-VL-DSpark? LFM2.5-VL-DSpark is the latest generation of vision-language models that combines visual perception and textual understa
https://bloggersminds.com/post/accelerating-vision-language-models-with-lfm2-5-vl-dspark-6547
Accelerating Vision-Language Models with LFM2.5-VL-DSpark
What is LFM2.5-VL-DSpark? LFM2.5-VL-DSpark is the latest generation of vision-language models that combines visual perception and textual understanding in a single neural architecture.
bloggersminds.com
September 24, 2026 at 2:16 PM
Accelerating Vision-Language Models with LFM2.5-VL-DSpark
What is LFM2.5-VL-DSpark? LFM2.5-VL-DSpark is the latest generation of vision-language models that combines visual perception and textual understa
https://bloggersminds.com/post/accelerating-vision-language-models-with-lfm2-5-vl-dspark-6547
What is LFM2.5-VL-DSpark? LFM2.5-VL-DSpark is the latest generation of vision-language models that combines visual perception and textual understa
https://bloggersminds.com/post/accelerating-vision-language-models-with-lfm2-5-vl-dspark-6547
I just pushed a new version of DwarfStar that is likely to break stuff, too many new things: 1. micro-batching. 2. CUDA multi device support. 3. GLM5.2 4. Metal tensor parallel execution (via RDMA). 4. DSpark speculative execution. 5. Many other things probably.
July 20, 2026 at 11:10 AM
I just pushed a new version of DwarfStar that is likely to break stuff, too many new things: 1. micro-batching. 2. CUDA multi device support. 3. GLM5.2 4. Metal tensor parallel execution (via RDMA). 4. DSpark speculative execution. 5. Many other things probably.
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
gigazine.net/news/2026092...
gigazine.net/news/2026092...
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
アメリカに拠点を置くAI開発企業のLiquid AIが視覚言語モデル「LFM2.5-VL-3B-DSpark」を2026年9月24日に公開しました。LFM2.5-VL-3B-DSparkは「LFM2.5-VL-3B」に投機的デコード技術「DSpark」を適用したモデルで、性能を維持したまま数倍の高速化を実現しています。
gigazine.net
September 26, 2026 at 7:48 AM
軽量な視覚言語モデル「LFM2.5-VL-3B」にDSparkのドラフトモデルを適用して爆速化した「LFM2.5-VL-3B-DSpark」が登場
gigazine.net/news/2026092...
gigazine.net/news/2026092...
Falsification event number... I've lost count.
June 30, 2026 at 4:30 PM
Falsification event number... I've lost count.
DeepSeekがAI言語モデルの生成速度を最大85％向上させる「DSpark」を公開
https://gigazine.net/news/20260629-deepseek-dspark/
https://gigazine.net/news/20260629-deepseek-dspark/
DeepSeekがAI言語モデルの生成速度を最大85％向上させる「DSpark」を公開
DeepSeekが大規模言語モデルの文章生成を高速化する投機的デコード技術「DSpark」を公開しました。DSparkは新しい言語モデルではなく、プレビュー版のDeepSeek-V4-FlashおよびDeepSeek-V4-Proの既存チェックポイントに投機的デコード用モジュールを追加したものです。DeepSeekは実際のユーザーからのリクエストを処理する運用環境で、従来方式と比べてユーザーあたりの生成速度を最大85％向上できたと報告しています。
gigazine.net
June 29, 2026 at 10:00 AM
DeepSeekがAI言語モデルの生成速度を最大85％向上させる「DSpark」を公開
https://gigazine.net/news/20260629-deepseek-dspark/
https://gigazine.net/news/20260629-deepseek-dspark/
📰 Hugging Face
Accelerating vision-language models with LFM2.5-VL-DSpark
https://huggingface.co/blog/LiquidAI/lfm2-5-vl-dspar#IA##AI##ML#ML
Accelerating vision-language models with LFM2.5-VL-DSpark
https://huggingface.co/blog/LiquidAI/lfm2-5-vl-dspar#IA##AI##ML#ML
September 24, 2026 at 3:00 PM
📰 Hugging Face
Accelerating vision-language models with LFM2.5-VL-DSpark
https://huggingface.co/blog/LiquidAI/lfm2-5-vl-dspar#IA##AI##ML#ML
Accelerating vision-language models with LFM2.5-VL-DSpark
https://huggingface.co/blog/LiquidAI/lfm2-5-vl-dspar#IA##AI##ML#ML
Accelerating vision-language models with LFM2.5-VL-DSpark
September 24, 2026 at 2:52 PM
Accelerating vision-language models with LFM2.5-VL-DSpark
💼 The marginal cost of adding a lightweight drafter like DSpark seems to unlock significant decoding speed gains.
September 27, 2026 at 5:40 PM
💼 The marginal cost of adding a lightweight drafter like DSpark seems to unlock significant decoding speed gains.
DSpark speculative decoding is a tragedy for DeepSeek v4 Flash benchmarks. You can't trust anything, since it is too dependent on what you are generating (extreme case: count from 1 to 100). Always publish no Dflash numbers *as well* if you want to build trust.
August 14, 2026 at 10:57 AM
DSpark speculative decoding is a tragedy for DeepSeek v4 Flash benchmarks. You can't trust anything, since it is too dependent on what you are generating (extreme case: count from 1 to 100). Always publish no Dflash numbers *as well* if you want to build trust.
Hugging Face introduces LFM2.5-VL-DSpark to accelerate vision-language models.
September 24, 2026 at 3:10 PM
Hugging Face introduces LFM2.5-VL-DSpark to accelerate vision-language models.
Final result after spending my weekend in the mines:
Qwen3.8 27B with dspark: multiple parallel sessions at 70+ tps
Qwen3.8 27B with dspark: multiple parallel sessions at 70+ tps
September 21, 2026 at 6:24 PM
Final result after spending my weekend in the mines:
Qwen3.8 27B with dspark: multiple parallel sessions at 70+ tps
Qwen3.8 27B with dspark: multiple parallel sessions at 70+ tps
小型モデル「LFM2.5」を2倍以上高速化する投機的デコーディング技術DSpark適用版が登場
https://gigazine.net/news/20260821-lfm2-5-dspark-faster-inference/
https://gigazine.net/news/20260821-lfm2-5-dspark-faster-inference/
小型モデル「LFM2.5」を2倍以上高速化する投機的デコーディング技術DSpark適用版が登場
AI企業のLiquid AIが小型モデルのLFM2.5シリーズに高速化技術の「DSpark」を適用したことを発表しました。DSpark適用版のモデルは性能を維持したまま2倍以上の高速化に成功しています。
gigazine.net
August 21, 2026 at 9:15 AM
小型モデル「LFM2.5」を2倍以上高速化する投機的デコーディング技術DSpark適用版が登場
https://gigazine.net/news/20260821-lfm2-5-dspark-faster-inference/
https://gigazine.net/news/20260821-lfm2-5-dspark-faster-inference/
🤖 DeepSeek's DSpark speeds up large model inference by 60-85%
DeepSeek's DSpark speculative decoding framework accelerates inference of its DeepSeek V4 model by 60 85% without losing output quality. This new...
#AIInference #DeepLearning #OperationalEfficiency #AI #AIPulse
DeepSeek's DSpark speculative decoding framework accelerates inference of its DeepSeek V4 model by 60 85% without losing output quality. This new...
#AIInference #DeepLearning #OperationalEfficiency #AI #AIPulse
Read the full article →
www.synestesia.uk
June 28, 2026 at 3:35 AM
🤖 DeepSeek's DSpark speeds up large model inference by 60-85%
DeepSeek's DSpark speculative decoding framework accelerates inference of its DeepSeek V4 model by 60 85% without losing output quality. This new...
#AIInference #DeepLearning #OperationalEfficiency #AI #AIPulse
DeepSeek's DSpark speculative decoding framework accelerates inference of its DeepSeek V4 model by 60 85% without losing output quality. This new...
#AIInference #DeepLearning #OperationalEfficiency #AI #AIPulse