as you wish👊
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…
0
0
1
93
visit NanoBench v0.5.0 here: shortbench.vercel.app
1/4 Small model. Strong agent result. 🚀
Our Spark-X2.6-4B-GGUF scored 94.6/100 in a community-reported NanoBench run with Hermes Agent: 9/12 tasks solved.
Just 4B parameters. Thank you, userr99, for testing and sharing the result! Screenshot below 👇
0
0
0
8
it's really great job, dario!
update on the @SparkLLM ced retrofit: v1 lost a lot of long context retrieval (nolima went from 33% to 8%). turned out the split at layer 18 was the problem, the later attention layers need stuff that just isnt in the layer 18 state yet. moved the split to layer 24 and it comes % for v1, stock is 59%
1
0
2
48
Tylerwtzhang retweeted
Proof that the future of incredible AI performance will live in a small model. @karpathy was right. I didn't believe it until I tried Spark.
Spark-X2.6-1.7B: less than half the parameters of our 4B, close on two benchmarks:
SWE Verified: 59.8 vs 62.8
τ²-bench: 85.3 vs 86.9
Both at high thinking.
Try it & Like on HF!
huggingface.co/XHToken/Spark-…
@huggingface @ollama @ModelScope2022
1
1
1
188
😃
@Tylerwtzhang @vaiduakhu @jaro3th @noctus91 Just tried a few and WOW. It seems like a qwen3.6-9b we never got. Its more efficient than 9b, wastes less token and get more things done. Its perfect por personal assistant like hermes. Currently getting 82tks median with a rx 9070xt and using it at q8.
0
0
2
102
sota on-device model in 2B class👉
Spark-X2.6-1.7B: less than half the parameters of our 4B, close on two benchmarks:
SWE Verified: 59.8 vs 62.8
τ²-bench: 85.3 vs 86.9
Both at high thinking.
Try it & Like on HF!
huggingface.co/XHToken/Spark-…
@huggingface @ollama @ModelScope2022
0
0
4
94
Tylerwtzhang retweeted
🎉 Spark-X2.6-4B reached #4 on Hugging Face Trending!
Snapshot: All · last 7 days, Oct 25, 2026.
Thanks for your support, hands-on reports & constructive feedback. If you find it useful, give it a ❤️ Like:
huggingface.co/XHToken/Spark-…
Snapshot: All · last 7 days, Oct 25, 2026.
Thanks for your support, hands-on reports & constructive feedback. If you find it useful, give it a ❤️ Like:
huggingface.co/XHToken/Spark-…
15
7
116
5,924
Tylerwtzhang retweeted
When should the cloud execute for the edge agent, and when should it just advise?
We present a systematic study of edge-cloud collaboration in long-horizon agent execution. Using EC²-Bench, we evaluate four collaboration strategies across 128 executable tasks, jointly measuring task completion, resource cost, and privacy exposure.
Here are three key findings:
1️⃣ Small cloud budget, big gains.(cost-efficiency). With Qwen3.5-9B on-device and GLM-5.3 in the cloud, edge-only scores 39.12; the best collaboration reaches 60.46 — about 73% of cloud-only (82.70) — at only $8.73, roughly 15% of cloud-only's $56.42.
2️⃣ Performance converges, privacy doesn't. On Qwen3.7-Plus, best collaboration hits 76.84 vs 82.70 cloud-only; strategy gaps shrink to just 3.67 completion points, yet privacy exposure spans 66.23 percentage points — and the best-performing strategy leaks the most.
3️⃣ The mode flips with edge capability. Weak edges do best with an external router that lets the cloud execute (2B: Per-step Routing 39.97 vs self-delegation 15.86); stronger edges do best when the cloud only advises (9B Cloud Advisory 60.46).
Further experiments highlight two abilities that matter for effective self-directed collaboration: knowing when to ask for help, and actually acting on the cloud's advice.
📄 Paper: arxiv.org/abs/2610.17816
🌐 Project: passionate11.github.io/ec2-bench-proj…
💻 Code: github.com/OpenBMB/EC2-Be…
#EdgeCloudCollaboration #AIAgent #LLM #OnDeviceAI #AgentBenchmark #EC2Bench #OpenBMB #Tsinghua
We present a systematic study of edge-cloud collaboration in long-horizon agent execution. Using EC²-Bench, we evaluate four collaboration strategies across 128 executable tasks, jointly measuring task completion, resource cost, and privacy exposure.
Here are three key findings:
1️⃣ Small cloud budget, big gains.(cost-efficiency). With Qwen3.5-9B on-device and GLM-5.3 in the cloud, edge-only scores 39.12; the best collaboration reaches 60.46 — about 73% of cloud-only (82.70) — at only $8.73, roughly 15% of cloud-only's $56.42.
2️⃣ Performance converges, privacy doesn't. On Qwen3.7-Plus, best collaboration hits 76.84 vs 82.70 cloud-only; strategy gaps shrink to just 3.67 completion points, yet privacy exposure spans 66.23 percentage points — and the best-performing strategy leaks the most.
3️⃣ The mode flips with edge capability. Weak edges do best with an external router that lets the cloud execute (2B: Per-step Routing 39.97 vs self-delegation 15.86); stronger edges do best when the cloud only advises (9B Cloud Advisory 60.46).
Further experiments highlight two abilities that matter for effective self-directed collaboration: knowing when to ask for help, and actually acting on the cloud's advice.
📄 Paper: arxiv.org/abs/2610.17816
🌐 Project: passionate11.github.io/ec2-bench-proj…
💻 Code: github.com/OpenBMB/EC2-Be…
#EdgeCloudCollaboration #AIAgent #LLM #OnDeviceAI #AgentBenchmark #EC2Bench #OpenBMB #Tsinghua
3
6
26
2,223
🚀 Spark-X2.6-4B is now #5 on Hugging Face’s model trending list! @huggingface
Thanks for all the support! 🤗 Give it a try and let us know what you think.
👉 Try huggingface.co/XHToken/Spark-…
#Spark #HuggingFace #OpenSourceAI
Thanks for all the support! 🤗 Give it a try and let us know what you think.
👉 Try huggingface.co/XHToken/Spark-…
#Spark #HuggingFace #OpenSourceAI
0
0
8
104
thanks. keep going…
A 4B model is matching 9B models on several benchmarks. Spark-X2.6-4B: same arch as X2.5, gains all from post-training — coding, agents, 1M context, on-device. The small-model curve is steep. reddit.com/r/LocalLLaMA/c…
0
0
0
34
Tylerwtzhang retweeted
@SparkLLM @huggingface @ModelScope2022 @ollama @vllm_project @sgl_project 🚨 Update🚨
We may have the best <16GB VRAM model over here. Impressive results and impressive speed.
⚠️The benchmark is about RAW performance. Thinking OFF, and MTP OFF.
Spark had difficult time with Terminal-Bench 2.0 (samples)
But Overall TOP #1 quality/speed for sure.
We may have the best <16GB VRAM model over here. Impressive results and impressive speed.
⚠️The benchmark is about RAW performance. Thinking OFF, and MTP OFF.
Spark had difficult time with Terminal-Bench 2.0 (samples)
But Overall TOP #1 quality/speed for sure.
0
1
3
51
Please tell us the performance on your own benchmarks
@NayteThaGreat Yeah going to test it myself on my own benchmarks. So far no small model has been able to beat Qwen3.5-9B. I’m getting ~70 tps so far on this new model
1
0
1
50
nice work!👉
@SparkLLM @huggingface @ModelScope2022 @ollama @vllm_project @sgl_project Wow made with Spark-X2.6-4B in less than 1 minute using 12k token only
0
0
0
30
If you find our model helpful, feel free to give us a star at👊:
github.com/XHToken/Spark-…
huggingface.co/XHToken/Spark-…
github.com/XHToken/Spark-…
huggingface.co/XHToken/Spark-…
Spark-X2.6-4B is here! Stronger coding, agents & long-context skills through improved post-training. Thinking: none / low / high.
@huggingface @ModelScope2022 @ollama @vllm_project @sgl_project
huggingface.co/XHToken/Spark-…
0
0
4
127
Spark-X2.6-4B retains exactly the same architecture as Spark-X2.5-4B, achieving its capability gains through improved post-training! 👉
Spark-X2.6-4B is here! Stronger coding, agents & long-context skills through improved post-training. Thinking: none / low / high.
@huggingface @ModelScope2022 @ollama @vllm_project @sgl_project
huggingface.co/XHToken/Spark-…
0
0
4
143
Spark-X2.6-4B / 1.7B will soon be open source. We’re excited to explore what’s possible with local AI, together with you. @SparkLLM
See you next week!
See you next week!
0
0
10
4,426
👉
--Spark-X2.5-4b
Stop what you're soing and make a local profile with the Spark-X2.5-4b asap.
for your @NousResearch hermes-agent.
mac-mini M2-pro 32gb ram
definitely the highest pass on the fields test of any 4B model I've ever used a few websites I generated with it in the tally smoothly. It understands that it's vision capable it ran for over an hour and 20 minutes on its own and stopped properly, validating its own work on some website updates very pleased with this model. I have not been having the same results with the 1.7 B model or even the same kind of feels but will continue to investigate.
0
0
1
98
👉
@SparkLLM I know everyone beelines for the agentic coding benchmarks, but anecdotally this is very smart. The 4b q8 quant passed a reasoning benchmark of mine that Qwen 3.5 9B Q4 did not, which is probably the strongest model I've used around that VRAM size. Good work.
0
0
0
27
👉
@Tylerwtzhang4@ModelScope2022-9B — intel benchmark (LocalBench)**
Custom for my own use case.
Both run on a Mac Mini M4 Pro, same agentic intel suite (multi-tool, tool-loop, structured Czech dreamer/nightly, deep reasoning). Intelligence only — not speed.
**Thinking mode (their natural mode):**
**multi_tool**
• Spark 4B: 85
• Ornith 9B: 85
**tool_loop**
• Spark 4B: 100
• Ornith 9B: 100
**dreamer**
• Spark 4B: 100
• Ornith 9B: 100
**nightly**
• Spark 4B: 100
• Ornith 9B: 100
**deep**
• Spark 4B: 75
• Ornith 9B: 70
****intel_score****
• Spark 4B: **96.2 ✅**
• Ornith 9B: **96.2 ✅**
Both pass the gate (≥85). Non-thinking mode fails for both (75.7 / 77.5) — these are reasoning models, they need thinking on.
0
0
0
37