SparkLLM (@SparkLLM) | nitter

6h

One RTX 3060. 12GB VRAM. 🔥

@DogukanUrker put Spark-X2.6-4B Q8_0 to the test:
✅ 288K context window
✅ Needle retrieved at 280K tokens

Thanks for sharing the results and config! 🙌 We’re working on even better models for future releases. Follow @SparkLLM to stay updated—we’d love to hear what you think when they’re out! 🚀

Doğukan
@DogukanUrker
Nov 2
running Spark-X2.6-4B on a single RTX 3060. 12GB VRAM. (config below)

288K context. ~67 tok/s decode, ~3500 tok/s prefill.

only 9 of its 36 layers keep a full kv cache, so long context is cheap. found a needle hidden at 280K tokens. Q8 over BF16: almost lossless at 4B, and s.

config:
llama-server
-m Spark-X2.6-4B-Q8_0.gguf
-ngl 99 -c 294912 -fa on --jinja
-np 1
--cache-type-k q8_0 --cache-type-v q8_0
-b 2048 -ub 1024
--temp 1.0 --top-p 0.95 --top-k 0 --min-p 0.0 --presence-penalty 0.0
--repeat-penalty 1.0
--reasoning on

want more context? swap to --cache-type-k q4_0 --cache-type-v q4_0 and -c 458752. 448K on the same card, needle still found at 436K. decode drops to ~12 tok/s that deep.

0

0

13

518

22h

Spark on phones: explore PocketPal, TokenSpark, LlamaPal and iPhone Canvas. Our Wiki collects 9 app cases with versions and sources. The PocketPal chart is a workstation web-search eval, not phone speed.
github.com/XHToken/Spark-…

1

0

9

612

22h

Replying to @SparkLLM

Android: TokenSpark offers an X2.5 trial; LlamaPal’s developer announced X2.5 4B support. On iPhone, Local LLM’s creator demonstrates X2.6 4B in Canvas. Creator images and speed claims are not independent device benchmarks.

1

0

1

71

22h

Replying to @SparkLLM

Also in the Wiki: AIでチャット, LocalEngine, Android Local-LLM, Weixuan and LocalAI. Evidence ranges from app listings to project support lists. Check the version and device limits, then share your own setup and results.
github.com/XHToken/Spark-…

0

0

2

60

Nov 2

Spark-X2.6-4B in two live web assistants: English questions, real replies.

Built something with Spark-X2.6? Share your demo in the replies!

Full case: huggingface.co/XHToken/Spark-…

3

0

17

1,093

23h

Replying to @SparkLLM

A candy-matching game built with Spark-X2.6-4B 🍬

Generated and refined in the Playground. Here is the real gameplay: match colors, score points, restart.

Build your own: awesome-astron-workflow.dev/playground

0

0

2

67

Nov 3

Six orders. Only four payments. Can Spark-X2.6 explain why?

Six orders in the database. Should all six count as revenue?

youtube.com/shorts/kdzDwZF…

1

0

8

397

Nov 3

Replying to @SparkLLM

One was refunded. Another was cancelled. We asked local Spark-X2.6 to write the SQL, then ran its exact query in SQLite.

The result: four payments, 5,749 cents—$57.49.

1

0

2

106

Nov 3

Replying to @SparkLLM

But how do we know the answer holds up? Watch us check it with an independent sum and a fresh database read.

Six fictional orders. One real local model call. A small test you can reproduce.

#SparkX26 #SQLite #LocalLLM #SQL #DataValidation

0

0

0

68

Nov 2

Replying to @SparkLLM

Both Spark-X2.6-4B demos in video format, as the original GIF isn't loading.

Built an app with Spark-X2.6? Share your demo in the replies!

Full case: huggingface.co/XHToken/Spark-…

0

0

1

114

Nov 2

Replying to @SparkLLM

Try the apps:
astronclaw-tutorial.space/en/
awesome-astron-workflow.dev/chat

Both offer None / Low / High thinking levels. Idle waiting is shortened where labeled; replies are unchanged.

Details: huggingface.co/XHToken/Spark-…

0

0

2

121

Nov 1

769 tok/s on a single RTX 4090 ⚡

Spark-X2.6-1.7B + DSpark
GSM8K · Q4_K_M · llama.cpp

Small model. Faster local inference.

❤️ Like & download on @huggingface + @ModelScope2022:
huggingface.co/XHToken/Spark-…
modelscope.cn/models/XHToken…

5

5

51

1,879

Nov 1

Replying to @SparkLLM

@huggingface @ModelScope2022 Just 345M draft parameters. Up to 7 draft tokens per pass. 🪄

Peak speedup: 2.43x (GSM8K, BF16, RTX 4090).
Average BF16 speedup across six datasets:
• 2.10x RTX 4090 / llama.cpp
• 1.88x M5 Pro / MLX

Full benchmark tables in the poster; results vary by workload.

2

0

4

256

Nov 1

Replying to @SparkLLM

@huggingface @ModelScope2022 Calling @sgl_project and the open inference community: try DSpark on your workloads!

BF16 GGUF draft + Spark-X2.6-1.7B target. Setup uses XHToken/llama.cpp; upstream integration is in progress.

❤️ Like the model, download & share your results:
huggingface.co/XHToken/Spark-…

0

0

4

181

Oct 31

DSpark for Spark-X2.6-4B is here! ✨

Average BF16 speedup:
2.14x RTX 4090 (llama.cpp)
2.50x M5 Pro (MLX)

Download & ❤️ Like on @huggingface and @ModelScope2022!
huggingface.co/XHToken/Spark-…
modelscope.cn/models/XHToken…

3

13

56

2,541

Oct 31

Replying to @SparkLLM

@huggingface @ModelScope2022 Faster inference. More magic. 🪄

624M BF16 draft, up to 7 draft tokens per pass.
Peak speedup: 2.94x on MATH500 (BF16, M5 Pro).
Peak throughput: 532 tok/s on GSM8K (Q4_K_M, RTX 4090).

Results vary by device, precision and workload; full tables in the model card.

2

0

3

304

Oct 31

Replying to @SparkLLM

@huggingface @ModelScope2022 Calling @sgl_project and the open inference community: try DSpark and share your results! ✨

The model card includes setup for our llama.cpp fork; upstream integration is in progress.

Like the model. Download the magic. ❤️
huggingface.co/XHToken/Spark-…

0

0

5

234

Oct 30

🎃 3.5M+ downloads. One magical community! ✨

Spark X2.5 + X2.6 have passed 3.5 million combined downloads. 🪄

This Halloween, create your own magic with Spark open models! 👻🍬

🤗 huggingface.co/XHToken
#Halloween #OpenSource

5

2

44

1,142

Oct 30

Replying to @SparkLLM

🧙‍♀️ Behind the magic: 204 community derivatives! ✨

Quantizations, fine-tunes & format conversions, made by builders like you. 🛠️

Thank you for every download, experiment and contribution. ❤️

Oct 30 snapshot; totals include pulls and approximate counts. 📊

1

0

2

191

Oct 30

Replying to @SparkLLM

🪄 Choose your Halloween build:
👻 A spooky story bot
🎃 A pumpkin-game NPC
🍬 A candy-themed coding helper

Explore Spark X2.6 in 1.7B / 4B:
huggingface.co/collections/XH…

Share your demo & give the models a ❤️ Like! ✨

0

0

1

285