Spawn Lau

iOS & Android indie dev behind CoolMeow & Cloakshot. I also experiment with on-device AI and share the numbers.

[Spawn Lau @SpawnLauv]

I build CoolMeow and Cloakshot.

CoolMeow lets friends log expenses together in multiple currencies and split bills quickly.
Cloakshot helps you review and hide sensitive details in screenshots.

I share product decisions and what I learn.

My apps: https://t.co/c6tth64swS

[Spawn Lau @SpawnLauv]

Gonna stress-test its basic math and see how it handles agent tasks on a phone.🥰

[SparkLLM @SparkLLM]

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!
https://t.co/kjB7OxCIZU
@huggingface @Ollama @ModelScope2022

[Spawn Lau @SpawnLauv]

@CodeWithSravya X Blue check...

[Spawn Lau @SpawnLauv]

@MiaAI_lab This harness can run on a phone with a LAN-hosted provider. https://t.co/YWRoJsqO0B

[Spawn Lau @SpawnLauv]

@AlfinCodes were these generated in one shot with no post-processing?

[Spawn Lau @SpawnLauv]

@exiztinz @itsmeMehak_ just use it to view code 😂

[Spawn Lau @SpawnLauv]

@itsmeMehak_ what's difference between js object, json, jsonl, jsonp ?

[Spawn Lau @SpawnLauv]

@vibeonX69

[Spawn Lau @SpawnLauv]

@codewith55 Slow like C ? C is slow? Assembly code?

[Spawn Lau @SpawnLauv]

@JayKranda why is the app rated 18+?🤣🤣🤣

[Spawn Lau @SpawnLauv]

@komal_uk01 ts

[Spawn Lau @SpawnLauv]

@Palak3312 where is the users, lol

[Spawn Lau @SpawnLauv]

@Tylerwtzhang I’m running model and agent tasks directly on my phone. A 1M-token context takes around ~35 GB, and the phone only has 12 GB total.

[Spawn Lau @SpawnLauv]

Tested several Spark-X2.6-4B variants. Q4_0 + OpenCL gave me the best overall performance.

Then I used it in BoltPocket to check the weather and build 2048. Neither passed my tests.

The main issues I hit:
1. Tool chains with 20+ steps ran into errors.
2. At 16K, context needed frequent compaction, and the generated code had errors.
3. One 2048 attempt failed after nearly 30 minutes.

[Spawn Lau @SpawnLauv]

@itsmeMehak_ {

[Spawn Lau @SpawnLauv]

@coldniko @mhelcig do you think Strata has a good chance of making it to mobile?

[Spawn Lau @SpawnLauv]

Tested three models in BoltPocket on my phone (llama.cpp + Pi Agent):

Bolt E4B: works, but slow.
Bonsai 2: tool calling doesn't seem to work reliably in this setup.
Spark-X2.6-4B: peg-native errors. I'm checking llama.cpp's parser.🥲If I can resolve this, Spark might be the best fit for my harness.

[Spawn Lau @SpawnLauv]

@NayteThaGreat Yeah, I saw that on my S25 too. The model runs, but heat, battery drain and memory use are real constraints.
But that game used a cloud model with a local harness, though.

[Spawn Lau @SpawnLauv]

Open-sourced my mobile LLM harness.

BoltPocket (built on Pi Agent) lets me test local models like Bonsai 2 2B right on my phone via llama.cpp.
Also supports OpenAI and OpenRouter.

Demo of it generating a game locally 👇