What if cross-chain transfers didn’t require trusting a central pool?
That’s the idea behind TunnelSwap.
🔐 Individual escrow on each chain.
🔑 One shared secret.
⏳ Time-based refund conditions.
🔍 Verifiable on-chain transactions.
The mechanism has been demonstrated across XRPL, Stellar and Hedera.
The product is still being developed, and an independent audit is planned.
We’re building in the open.
Explore the technology 👇
paydlux.com/tunnel/?v=2
Paydlux is building.
We’re developing payment technology across blockchain ecosystems, experimenting with new ways to connect networks and creating infrastructure for the next generation of digital payments.
We don’t just want to talk about innovation.
We want to build it.
Our latest project is TunnelSwap.
Take a look at what we’re creating:
🌐 paydlux.com/tunnel/?v=2
Follow @paydlux for what’s coming next.
Replying to @CoinbaseMarkets
Legal stuff: You WILL loose your money. We are selling higher to make it seem like there is high demand on the stock, but every one knows we will dump on you.
There ya go I fixed it for ya!!
One blockchain shouldn’t define the limits of payments.
🌐 Stellar.
⚡ XRPL.
🔷 Hedera.
🚀 And the wider blockchain ecosystem.
Paydlux is actively building and experimenting across networks.
Different chains. Different capabilities. One ambition: more connected payment infrastructure.
The future is interconnected.
The XRP Ledger is an important part of the blockchain payment landscape.
At Paydlux, we’re exploring and developing payment solutions that leverage the capabilities of different networks.
Our TunnelSwap mechanism has demonstrated conditional escrow transfers involving XRP and XLM.
We believe interoperability deserves more attention.
Building on Stellar.
We believe blockchain technology should solve real problems.
That’s why we’re developing payment infrastructure with real-world applications in mind.
Less friction. More possibilities. Better payment experiences.
And we’re only getting started.
The future of payments isn’t built on just one blockchain.
At Paydlux, we’re actively building, integrating and innovating across blockchain ecosystems.
Stellar. XRPL. Hedera. And beyond.
Our mission? Make blockchain payments more accessible, connected and useful in the real world.
We’re just getting started.
Developers: AI agents need more than intelligence.
They need infrastructure.
Think about what happens when agents start buying:
🔹 API calls
🔹 Premium datasets
🔹 Computing resources
🔹 Digital services
🔹 Specialized tools
The next generation of AI applications needs a payment layer built for machines.
Paydlux is building exactly around this challenge:
• Agent-compatible 402 payment flow
• USDC settlement on Stellar
• Ledger-enforced spending mandates
• Human-controlled payment release
And developers can get started with:
npm install @paydlux/agent
Open source. Built for the agent economy.
Explore the code and get involved:
paydlux.com
#BuildInPublic #Developers #AIAgents #Stellar
What if your AI agent could request a payment without ever controlling the release of your money?
That’s the idea behind Paydlux. 🚀
Here’s the flow:
1️⃣ Agent requests a purchase.
2️⃣ Paydlux checks its spending mandate.
3️⃣ Funds are held securely.
4️⃣ You approve or deny from your phone.
5️⃣ The seller gets paid after approval.
Built on Stellar. Settled in USDC.
The agent makes the request. YOU make the decision.
Explore: paydlux.com
#USDC #Stellar #AIAgents #Web3
Z-B retweeted
API
Model id - `qwen3.8-flash-next-iq3_xxs` |
Context | 131072 tokens
Control | 0.1.59.1
Engine | 0.1.59, source `f-0213-rc 712fec6b4297`, CUDA 13.0, portable |
GPU | NVIDIA GeForce RTX 4090 Laptop GPU, compute capability 8.9, 16376 MiB |
| Driver | 615.78.08 |
| `CUDA_VISIBLE_DEVICES` | `0` |
| `CUDA_DEVICE_ORDER` | `PCI_BUS_ID` |
Process tree
| 671647 | launcher | `strata.py launch` |
| 672937 | runtime | `python3.12 -m strata_control.runtime --host launch-iq3_xxs-128k --shutdown-timeout 30 --startup-wait 886` |
| 672944 | API server | `python3.12 -m serve.server --engine strata --config ~/.local/state/strata/launch/iq3_xxs-128k.json --host 127.0.0.1 --port 8080 --shutdown-timeout 30.0` |
| 673050 | engine | the argv in the next section |
Server cwd: `~/.local/share/strata/components/control/0.1.59.1`
Engine RSS while up: about 44.6 GiB. GPU memory used by the engine: 14514 MiB.
## Engine argv (what was actually running)
Saved tuning `IQ3_XXS 128K` added `--conversation-cache-mib 3072 --max-context 131072 --vram-reserve-mib 704`.
```
strata --serve \
--pack /home/****/models/packs/iq3_xxs \
--native /home/****/models/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf \
--ple-gguf /home/*****/models/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf \
--expert-profile /home/***/.local/share/strata/components/control/0.1.59.1/data/expert-profile.bin \
--expert-cache auto \
--prefill auto \
--kv int8 \
--kv-resident 32768 \
--pool-workers 15 \
--conversation-cache-mib 3072 \
--max-context 131072 \
--vram-reserve-mib 704 \
--draft dflash2 \
--draft-model /home/***/models/dflash2/qwen3.8-flash-next-dflash2.q4_0-selq4.pack \
--draft-fallback none \
--mtp-draft-vocab /home/****/models/mtp/rt/draft_vocab.bin \
--spec 6 \
--draft-conf 0.7
```
| Flag | Value |
|---|---|
| `--pack` | `/home/****/models/packs/iq3_xxs` |
| `--native` | `Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf` (43.81 GiB) |
| `--ple-gguf` | `Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf` (26.82 GiB) |
| `--expert-profile` | control `data/expert-profile.bin` (24576 ranked pairs, 24576 slots) |
| `--expert-cache` | `auto` → 5017 slots, 8.14 GiB VRAM, profile policy, no eviction |
| `--prefill` | `auto` → prompt chunks of 8192 tokens, 384-slot ring |
| `--kv` | `int8` |
| `--kv-resident` | `32768` of 131072 cells per QSA layer in VRAM; K/V also in 1.55 GiB pinned RAM |
| `--pool-workers` | `15` (logical CPUs 2,4,6,8,10,12,14,16–23; host thread on CPU 0) |
| `--conversation-cache-mib` | `3072` |
| `--max-context` | `131072` |
| `--vram-reserve-mib` | `704` (plus 346 MiB draft head, plus 256 MiB after the cache) |
| `--draft` | `dflash2` |
| `--draft-model` | `qwen3.8-flash-next-dflash2.q4_0-selq4.pack` (324 MiB file; 591 MiB VRAM, Q4_0) |
| `--draft-fallback` | `none` (no MTP layer loaded) |
| `--mtp-draft-vocab` | `/home/*****/models/mtp/rt/draft_vocab.bin` |
| `--spec` | `6` |
| `--draft-conf` | `0.7` |
## Resolved at startup
| Item | Value |
|---|---|
| Weights loaded | 39.97 GiB at 4.57 GiB/s |
| Pack weights | 1466 MiB from the pack; 302 canonical tensors served natively |
| Native projections | 300 matrices, 1781.11 MiB |
| Token embedding | IQ3_S, 260 MiB, mapped in host memory |
| Output head | experimental native Q5_K, 437043200 bytes |
| PLE | on, 320001536 rows, from shard 2 |
| Expert arena | file cache, `cudaHostRegister PORTABLE`; THP always; no hugetlb (`vm.nr_hugepages=0`) |
| CPU expert kernels | AVX-2 |
| CPU pool | 48 tasks/phase, 16 participating threads |
| Prefill CPU share | on for prompt chunks up to 3072 tokens |
| Prompt path | borrows 2566 CUDA0 cache slots (4.12 GiB) |
| PCIe probe | 12.8 GB/s host→device; serving `pcie_frac` 0.15 |
| Route tail skip | `STRATA_ROUTE_TAIL_SKIP=7` (default, on) |
| Adaptive overlap | off |
| Adaptive tier | asynchronous tier stayed on |
| Verify windows | captured 1 through 8 tokens; DFlash2 windows up to 8 |
| Draft head | 106299-token draft head, 178.4 MiB; selector codebooks Q4_0 |
| Hyper-connection read | staged (`cp.async`) |
| QSA prefill | norm + RoPE in one pass on; attention gate in the epilogue off |
| SSD keepalive | on (one page after 100 ms idle, until 60 s after the last request) |
| Sampling default | greedy when a request omits `temperature` |
| Vision | off |
| Tune | skip |
| Memory guard at start | limit 62.5 GiB, floor 2560 MiB |
## Model shape (from the pack)
| | |
|---|---|
| Family | qwen |
| Quant label | IQ3_XXS |
| Pack tool | `tools/iq_pack.py`, schema 1, pack index v3 |
| Layers | 48 (`blk.0`–`blk.47`) |
| Experts | 512 per layer, native-expert bytes 42912972800 |
| Hidden / embedding | 2560 |
| Vocab | 248320 |
| Tokenizer | `gpt2` / pre `qwen35`, no BOS |
| EOS | 248046 |
| BOS / pad | 248044 |
| Pack tensors | 1079 (align 64, pool 1538035200) |
| Tokenizer path | `/home/****/models/packs/iq3_xxs/tokenizer` |
## Launch record
File: `~/.local/state/strata/launch/iq3_xxs-128k.json`
| Field | Value |
|---|---|
| `model_name` | `qwen3.8-flash-next-iq3_xxs` |
| `host` / `port` | `127.0.0.1` / `8080` |
| `gpu` | `0` |
| `drafting` | `dflash2` |
| `open_browser` | false |
| `exe` | `component:strata-engine` |
| `tokenizer` | `/home/****/models/packs/iq3_xxs/tokenizer` |
| `engine_ready_s` | 946 |
| log | `~/.local/state/strata/launch/iq3_xxs-128k-hermes.engine.log` |
Choices file `iq3_xxs-128k.choices.json`: access `local`, context `131072`, drafting `dflash2`, family `qwen`, kv `int8`, model `IQ3_XXS`, port `8080`, tune `skip`, vision `off`.
Ready file said `api_state=ready`, `model_state=ready`, `model_id=qwen3.8-flash-next-iq3_xxs`, pid 672944.
Imagine this:
You give an AI agent a $10 budget.
A bad prompt, compromised tool or unexpected action makes it attempt a $1,000 purchase.
What stops it?
A limit written inside the agent’s code might not be enough.
That’s why Paydlux takes a different approach:
🛡️ Spending mandates enforced on the Stellar ledger.
📱 Human approval before payment release.
💰 Defined spending limits.
AI should be autonomous. Your money shouldn’t be unprotected.
paydlux.com
#AgentSecurity #AIInfrastructure #Stellar
Replying to @stoneiland
My analysis, we will visit $0.75. If that Zone doesn't hold we can call it "It's over".
Sad times










