InternetException

Just couple of days ago poolside released their new model: Laguna S 2.1. I was happy to try it since it have native support for 1M tokens in context. I am doing development of hobby projects on my phone, via SSH prompting models to do stuff and 250k tokens of context using Qwen 3.6 27B usually means going through few rounds of compaction before I am happy with the results. Having longer context would be better – in theory – without those ugly interruptions and reiterating stuff to agent that were lost in previous session.

So what is nice:

• 1M context without any configuration
• Model is quite capable with development tasks
• It MOE model so it is quite fast without MTP 20-10t/s (I tested it to around 500k context) on Strix Halo, for model this size

What is worse:

• It is MOE so it loops very often
• with longer context it loops more often
• this is why I was not able to test it till 1M tokens (stop, ask a question, see it looping again, rewind, ask different question etc.) at some point it felt like it is not progressing with the task at all
• it have MTP with drafter model but it is not supported yet in mainline llama.cpp, so I did not tested it yet
• official documentation states to use: temperature: 0.7, top_p: 0.95 for better results in longer contexts but that disabled thinking and model kept looping
• longer contexts restart is really painful since prefill drops to 90t/s for 300k tokens, so if you have 500k context restarting it means 1h-90min of waiting

Right now I am back to Qwen 3.6. Going back felt good because Laguna task was immediately picked up by Qwen and finished it in few minutes.

Summary

Laguna seems to be capable model, but it feels less capable for agentic tasks, and without close lookup, or strong guardrails will keep eating context looping around complex problems. 1M context would be nice but in overall does not seem to be worth it.

I bought new phone and I started installing applications from previous device. When I installed termux I tried to copy ssh keys and ssh config from previous installation.

This is where problems started. I could not reach any local address. Nothing in 10.x.x.x was working. No pings no ssh.

I could do that to public IPs just fine.

I tried to install new version of packages but it also did not work.

Termux already was at latest version, because it was just installed. Permissions for full network access was also fine.

I installed Graphene OS for first time and I tried to search anything relevant to similar issues regarding Termux on this system.

Nothing. Only results were some firewall settings on linux.

I though that maybe new version of termux is buggy somehow so I compared old phone termux version and on new phone.

Old phone had F-droid version when on new phone I installed it from Play store via Aurora.

Then I remembered having similar issues before, year or two ago. Seems like Play Store version is somehow buggy or changed to adhere some Google rules.

Anyway if you will not able to connect to local address via SSH or you will see something like:

ping: (sendmesg) operation not permitted

Install termux from f-droid instead.

If this will be still possible after 30th of September 2026.

I remember when I started selfhosting my own instance of GitLab, which had simple code editor that allows you to edit code quickly inline. Then they integrated with fork of Visual Studio Code and allowed you to code in real repository directly in GL.

I was thinking of doing quick code changes to applications at this way, change some configuration of my API, quick fix of my mobile phone for home automation and etc.

But it is hard!

It is hard to do anything substantial this way. You lack code completion, hints, terminal, debugger, tests and everything IDE gives you.

It was trial and error basically. And on mobile phone editing of files was terrible!

But now you have coding agents. You can just ask your PC to write some code and test it. You can ask it to do review and fix issues.

You can ask it to do design for you.

And if fails still very often doing silly stuff…

But it is thousands times better and you do not have to fix everything yourself. Of course there are other issues like hacked packages and malicious code that steals your credentials.

But having an ability to ask your PC to fix your tool that you made to send you emails about failed cronjob when you are on the bus… that is something new.

I am doing a lot of that via SSHing into one of my servers from my phone:

This is great!

This gives you that great felling of tinkering and achieving stuff without all of those problems, like broken syntax and debugging issue.

The only thing better than that would be probably ability to talk to my agent and explain what to do while relaxing or lawn mowing.

I do not know… I find it funny.

After few days fighting with local LLM to force it to write documentation and comments in code, in good quality I must add, and the forcing it to write readable unit tests, after I wrote:

File is untracked (new). Build passes. Ready for review.

I just answered.

awesome mate. It is friday we deserve a brake. High five.

I was thinking for a bit and answered:

I don’t know it gave me a chuckle.

Just out of curiosity I tested Qwen3.6-27B model with MTP speculative decoding with 6 draft tokens. Whole script looked like this:

docker run --rm --name="${2:-qwen3.6-medium-vllm}" \
    --group-add=video \
    --cap-add=SYS_PTRACE \
    --security-opt seccomp=unconfined \
    --device /dev/kfd \
    --device /dev/dri \
    -v /cache/huggingface:/root/.cache/huggingface \
    -v /cache/vllm:/root/.cache/vllm \
    --env "HF_TOKEN=$3" \
    -e TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 \
    -e VLLM_ROCM_USE_AITER=1 \
    -e FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE \
    -p "${1:-8000}":8000 \
    --ipc=host \
    --entrypoint '/bin/sh' \
    vllm/vllm-openai-rocm:nightly \
    -c "pip install fastokens; vllm serve Lorbus/Qwen3.6-27B-int4-AutoRound \
    --gpu-memory-utilization 0.48 \
    --dtype half \
    --optimization-level 1 \
    --enable-prefix-caching \
    --performance-mode interactivity \
    --kv-cache-dtype fp8_e4m3 \
    --language-model-only \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --speculative-config '{ \"method\": \"mtp\", \"num_speculative_tokens\": 2}' \
    --max-num-seqs 1 \
    --tokenizer-mode fastokens \
    --max-num-batched-tokens 262144 \
    #--max-model-len 262144"

Those are results with vLLM bench tool:

============ Serving Benchmark Result ============
Successful requests:                     100       
Failed requests:                         0         
Maximum request concurrency:             1         
Benchmark duration (s):                  3009.77   
Total input tokens:                      2915      
Total generated tokens:                  25600     
Request throughput (req/s):              0.03      
Output token throughput (tok/s):         8.51      
Peak output token throughput (tok/s):    4.00      
Peak concurrent requests:                2.00      
Total token throughput (tok/s):          9.47      
---------------Time to First Token----------------
Mean TTFT (ms):                          808.65    
Median TTFT (ms):                        819.95    
P99 TTFT (ms):                           842.02    
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          114.83    
Median TPOT (ms):                        110.50    
P99 TPOT (ms):                           145.64    
---------------Inter-token Latency----------------
Mean ITL (ms):                           400.91    
Median ITL (ms):                         407.92    
P99 ITL (ms):                            411.89    
==================================================

Not much better, despite GGUF MTP model having the same settings and being 6 times faster.