A production-stable deployment of AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4 with DFlash speculative decoding on NVIDIA DGX Spark (GB10 / sm_121a).
⚠️ READ THE REQUIREMENTS SECTION FIRST. This image and its weights are tuned specifically for the DGX Spark (GB10 / sm_120-121 Blackwell). It will NOT work on Hopper, Ampere, B200, or other Blackwell variants without rebuilding.
| Model | AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4 (~22 GB, multimodal — image input working as of the 2026-06-18 vision fix) |
| Drafter | z-lab/Qwen3.6-35B-A3B-DFlash (~905 MB, public anonymous pull) |
| Hardware | DGX Spark (NVIDIA GB10, 128 GB unified memory, sm_121a) |
| Image | ghcr.io/aeon-7/aeon-vllm-ultimate:latest (= :2026-06-18-v0.23.0-dflashfix; rollback :2026-06-11-pr41703) |
Complete recipe — pull the container, the NVFP4 weights, and a fresh DFlash drafter, then serve with the vetted DGX-Spark flags. (No --mamba-block-size for this 35B-A3B drafter — it is all-full-attention.) Run on a DGX Spark (GB10 / sm_121a); see Hard Requirements first.
# 1. Pull the unified vLLM container (anonymous pull)
docker pull ghcr.io/aeon-7/aeon-vllm-ultimate:latest
# 2. Pull the NVFP4 model weights (this repo's weights)
huggingface-cli download AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4 --local-dir ./aeon-model
# 3. Pull the DFlash drafter — FRESH (re-pull if cloned before 2026-04-19)
huggingface-cli download z-lab/Qwen3.6-35B-A3B-DFlash --local-dir ./aeon-drafter
# 4. Serve (ENTRYPOINT is /bin/bash, so pass --entrypoint vllm then serve ...)
docker run --gpus all --rm -p 8000:8000 \
-v ./aeon-model:/model:ro \
-v ./aeon-drafter:/drafter:ro \
--entrypoint vllm \
ghcr.io/aeon-7/aeon-vllm-ultimate:latest serve /model \
--served-model-name qwen36-35b-heretic qwen36-fast qwen36-deep \
--host 0.0.0.0 --port 8000 \
--quantization compressed-tensors \
--max-model-len 262144 \
--max-num-seqs 64 \
--max-num-batched-tokens 16384 \
--gpu-memory-utilization 0.85 \
--enable-chunked-prefill \
--enable-prefix-caching \
--trust-remote-code \
--limit-mm-per-prompt '{"image":4,"video":2}' \
--enable-auto-tool-choice --tool-call-parser qwen3_coder \
--reasoning-parser qwen3 \
--attention-backend flash_attn \
--speculative-config '{"method":"dflash","model":"/drafter","num_speculative_tokens":11}'Optional — faster restarts: add
-v ./vllm-cache:/root/.cache/vllmto thedocker run(after a one-timemkdir -p ./vllm-cache) to persist the FlashInfer NVFP4 GEMM autotuner + CUDA-graph capture across container restarts. First boot stays ~10–12 min, but subsequent restarts drop to ~3–5 min.
For a docker-compose deployment, the 3-alias setup, and per-flag rationale, see Deployment (compose + full recipe) below.
Measured on the current production image ghcr.io/aeon-7/aeon-vllm-ultimate:latest (vLLM 0.23.0 + AEON sm_121a build + DFlash num_speculative_tokens: 11), single-stream (c=1), by category. Mixed-domain prompt set, enable_thinking=false for clean decode-rate measurement:
| Category | 🟢 Decode tok/s | TTFT p50 | TPOT p50 | Prefill (PP) | DFlash accept |
|---|---|---|---|---|---|
| Coding | 91.7 | 88 ms | 10.9 ms | 509 tok/s | 33% |
| Math | 123.6 | 113 ms | 8.1 ms | 494 tok/s | 48% |
| Reasoning | 120.6 | 120 ms | 8.3 ms | 359 tok/s | 46% |
| Prose | 75.2 | 137 ms | 13.3 ms | 234 tok/s | 24% |
| Natural language | 91.8 | 104 ms | 10.9 ms | 326 tok/s | 32% |
| Extraction / JSON | 79.8 | 103 ms | 12.5 ms | 468 tok/s | 28% |
Single-stream decode averages ~97 tok/s across the six categories (min 75 on open-ended Prose, max 124 on Math/Reasoning where DFlash acceptance is highest). This A3B MoE is fast: math/code/reasoning prompts hit 120+ tok/s single-stream because the DFlash drafter accepts 46–48% of its proposals; open-ended Prose/Extraction settle around 75–92 tok/s at lower acceptance.
Stock baseline pending. There is no stock vanilla-vLLM baseline for this model yet — a fresh fully-vanilla benchmark (no DFlash, no AEON/sm_121a optimizations) on the current version is still to be run, so we do not quote a stock-vs-optimized speedup multiple here. The figures above are measured on the optimized
aeon-vllm-ultimate:latest(vLLM 0.23.0) build. The DGX-Spark win this image delivers is the c=64 concurrency hold (the prior image crashed at c≥32 under speculative decoding) plus long-context draft-acceptance hold — see What we fixed for the DGX Spark.
Aggregate throughput climbs cleanly to c=64 with zero crashes — the exact regime where the prior image (vllm-spark-omni-q36:v1.2) died at c≥32 under speculative decoding. Peak aggregate is ~740 tok/s at c=64 (Math), ~717 (Extraction/JSON), ~669 (Reasoning); throughput is DFlash-acceptance-driven, so creative-text categories run lower (~430 Prose / ~474 Natural-language at c=64, where draft acceptance is ~22–27% vs ~46–55% on structured text). Per-active-stream decode and TPOT degrade gracefully as concurrency rises.
| Concurrency | Math agg tok/s | Coding agg tok/s | Extraction/JSON agg tok/s | Best per-req decode p50 |
|---|---|---|---|---|
| 1 | 117.6 | 89.1 | 77.6 | 123.6 |
| 8 | 415.6 | 360.3 | 398.5 | 61.7 |
| 16 | 558.6 | 463.7 | 557.2 | 39.8 |
| 32 | 715.1 | 568.5 | 684.5 | 25.1 |
| 64 | 739.8 | 608.7 | 716.2 | 16.2 |
Best concurrency for chat UX: 8–16 (TTFT < 350 ms, per-req 25–62 tok/s); best for max aggregate throughput: 32–64. DFlash spec-decode acceptance is stable across concurrency (~24–48% position-0 depending on prompt class). Stress-tested with 32K-token prompts: long-context acceptance holds (see below).
Reconciliation note: earlier revisions of this card quoted 83.9 tok/s median single-stream and a ~313 tok/s aggregate plateau (and an
~785 tok/sfigure in the compose comments), all measured on the oldervllm-spark-omni-q36:v1.2image with DFlash k=15 and a different prompt mix/harness. Those are superseded by the per-category v0.23.0 figures above (≈97 tok/s avg single-stream, ~740 tok/s peak aggregate at c=64) measured onaeon-vllm-ultimate:latestat n=11.
Measured single-stream (c=1) at realistic agent-history depths on the same image:
| Context (measured prompt) | Coding decode | Reasoning decode | Extraction/JSON decode | DFlash accept (best) |
|---|---|---|---|---|
| ~16k tokens | 90.8 tok/s | 95.9 tok/s | 106.1 tok/s | ~52% |
| ~32k tokens | 79.3 tok/s | 73.0 tok/s | 94.4 tok/s | ~58% |
Draft acceptance does not collapse as context grows — at 32k it holds 42–58% across categories. (Unlike the 27B drafter, the z-lab Qwen3.6-35B-A3B DFlash drafter is an 8-layer all-full-attention model, so it never had a sliding-window collapse to begin with — the long-context hold here is intrinsic to the drafter, and the prefix-cache-safe DFlash path (PR #41703) keeps --enable-prefix-caching corruption-immune.) Prefill stays fast (~4.3–5.1k tok/s) so TTFT at 32k is single-digit seconds at c=1.
Re-validated 2026-06-19 on the exact published quickstart recipe (fresh
aeon-vllm-ultimate:latestpull + corrected weights, DFlashn=11): multimodal vision works end-to-end (7/7 on an image probe, 0 vision skip-loads) and the throughput spread reproduced with zero errors at c=64 — peak ~800 tok/s @ c=32 (Reasoning), structured categories 705–781 tok/s @ c=64, creative-text 430–474 @ c=64; DFlash acceptance ~46–55% (structured) / ~22–27% (creative) at short context, holding ~40–45% at 16–20K tokens.
Full bench results on the HF model card: AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4.
| Component | Required | Notes |
|---|---|---|
| GPU | NVIDIA GB10 (DGX Spark only) | sm_120 / sm_121a Blackwell. Other GPUs WILL NOT WORK with the published image. |
| Unified memory | 128 GB | Spark default |
| Disk | 35 GB free | Image (~22 GB) + weights (~22 GB) + drafter (~1 GB) + headroom |
Image will NOT work on:
- H100/H200 (sm_90 — Hopper)
- A100/A40 (sm_80 — Ampere)
- B200/GB200 (sm_100 — different Blackwell variant; rebuild from source)
- L40S/RTX 4090/RTX PRO 6000 (sm_89/sm_120 desktop variants — see
docs/build.md)
| Component | Version | Notes |
|---|---|---|
| NVIDIA driver | ≥ 580.x | nvidia-smi should print "NVIDIA GB10" |
| Docker | ≥ 25.x | with nvidia-container-toolkit |
| OS | Ubuntu 24.04 LTS confirmed | other Linux distros likely fine |
The DFlash drafter z-lab/Qwen3.6-35B-A3B-DFlash is now a public HF repo —
just hf download it directly, no token needed.
⚠️ If you cloned the drafter before 2026-04-19, you MUST re-pull. The earlier drafter had a long-context bug that causedcudaErrorIllegalAddresscrashes after ~16K tokens. The fixed version is on HF as of 2026-04-19.
For a one-block copy-paste that pulls everything and serves, see Quickstart (copy-paste) at the top. This section covers the docker-compose flow, the canonical on-disk layout, and the per-flag rationale.
# 1. Pre-flight check — confirm anonymous pull works
docker pull ghcr.io/aeon-7/aeon-vllm-ultimate:latest
# 2. Pull both models into the canonical layout
sudo mkdir -p /opt/qwen36 && sudo chown $USER:$USER /opt/qwen36
cd /opt/qwen36
export HF_HUB_ENABLE_HF_TRANSFER=1
hf download AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4 --local-dir ./qwen36-nvfp4 &
hf download z-lab/Qwen3.6-35B-A3B-DFlash --local-dir ./qwen36-dflash &
wait
# 3. Get the compose file
curl -fsSL https://raw.githubusercontent.com/AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4-DFlash/main/examples/docker-compose.yml \
-o docker-compose.yml
# 4. Start the server (3-5 min to first "Application startup complete")
docker compose up -d
docker compose logs -f
# 5. Smoke test (use temperature=0 for greedy → max DFlash speedup)
curl http://localhost:8000/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{
"model": "qwen36-fast",
"messages": [{"role":"user","content":"What is 17 × 23?"}],
"max_tokens": 2048,
"temperature": 0
}'The unified image's ENTRYPOINT is /bin/bash, so a raw docker run must pass --entrypoint vllm then serve ... (compose uses entrypoint: vllm + command: serve ...). The validated production flags for this model on DGX Spark:
vllm serve /models/qwen36 \
--served-model-name qwen36-35b-heretic qwen36-fast qwen36-deep \
--host 0.0.0.0 --port 8000 \
--quantization compressed-tensors \
--max-model-len 262144 \
--max-num-seqs 64 \
--max-num-batched-tokens 16384 \
--gpu-memory-utilization 0.85 \
--enable-chunked-prefill \
--enable-prefix-caching \
--trust-remote-code \
--limit-mm-per-prompt '{"image":4,"video":2}' \
--enable-auto-tool-choice --tool-call-parser qwen3_coder \
--reasoning-parser qwen3 \
--attention-backend flash_attn \
--speculative-config '{"method":"dflash","model":"/models/qwen36-dflash","num_speculative_tokens":11}'Key changes from the old vllm-spark-omni-q36:v1.2 recipe:
| Flag | Now | Was | Why |
|---|---|---|---|
| Image | aeon-vllm-ultimate:latest |
vllm-spark-omni-q36:v1.2 |
One unified vLLM 0.23.0 sm_121a image; the prior image crashed at c≥32 under DFlash |
--quantization |
compressed-tensors |
compressed-tensors |
unchanged (llm-compressor NVFP4) |
--attention-backend |
flash_attn |
flash_attn |
unchanged (body backend) |
DFlash num_speculative_tokens |
11 | 15 | n≈10–11 is optimal here — acceptance/long-context fidelity drop as n climbs past ~11; n=15 wasted draft compute |
| Drafter attention backend | default (do not set) | — | the non-causal DFlash drafter requires its default backend / BF16 KV; do not add attention_backend to the spec config or set --kv-cache-dtype |
--gpu-memory-utilization |
0.85 (LLM-only) · 0.70 (co-located) | 0.85 | 0.85 maximizes KV cache when the LLM is the only GPU service and is safe on GB10 (below the 0.88 cap); drop to ≤0.70 only if co-locating ASR/TTS/embeddings (page-thrash headroom) |
--max-num-batched-tokens |
16384 | 65536 | 65536 exceeds the compile-range ceiling (compile_ranges_endpoints: [32768]) → eager-prefill fallback + heavy unified-memory pressure at load. 16384 stays in the compiled range, frees several GB of prefill activation, and is throughput-neutral — chunked prefill preserves the full 256k context. |
--mamba-block-size |
(not set) | — | this drafter's attention is all-full-attention; no mamba-block-size override is needed |
The DFlash drafter is a no-regression migration onto the unified image — the 8-layer all-full-attention drafter behaves the same as on the old image at every concurrency (no long-context collapse to undo and nothing to gain past n≈11), so the win here is the unified container + the c≥32 concurrency-crash fix, not a raw single-stream speedup.
If max_tokens < ~1500 your response may show content: null with finish_reason: "length" — that's the model hitting max-tokens during reasoning, not a crash. See docs/troubleshooting.md. Use ≥ 2048 for thinking-enabled requests.
For the full step-by-step (with pre-flight + post-deploy verification), see docs/dgx-spark-setup.md.
The production container is now the unified ghcr.io/aeon-7/aeon-vllm-ultimate:latest (= :2026-06-18-v0.23.0-dflashfix; rollback :2026-06-11-pr41703) — vLLM v0.23.0 built from source for GB10 / sm_121a and merged with the AEON speculative-decoding stack. It supersedes the older vllm-spark-omni-q36 (v1/v1.2/v2) lineage of per-revision containers; see What we fixed for the DGX Spark for the current fix set.
The historical v1/v1.2 per-revision backport patch table is kept below for operators still on the old image. Each entry came from a real deployment failure, but several are now upstreamed or obsolete on the unified v0.23.0 base.
| # | Patch | What it fixes |
|---|---|---|
| 1 | register_qwen3_5_text.py |
Adds text-only Qwen3_5MoeForCausalLM to vLLM model registry. Legacy for v2 multimodal weights because they load through the canonical multimodal class, but harmless/backward-compatible for v1 text-layout weights. |
| 2 | patch_cuda_optional_import.py |
Wraps import vllm._C_stable_libtorch in RTLD_LAZY dlopen. Needed on older sm_120 builds where _C_stable_libtorch references unresolved SM100-only MXFP4 symbols. Newer vLLM builds may already export the needed stubs and can skip this. |
| 3 | patch_kv_cache_utils.py (×4 sites) |
Mamba/linear-attention groups could expose block_size=None to downstream arithmetic. Newer vLLM commits derive/validate mamba_block_size before these paths execute, so this is mainly an older-base backport. |
| 4 | patch_mrope_text_fallback.py |
Qwen3.6 declares M-RoPE in config but no model class implements get_mrope_input_positions in vLLM HEAD. Adds inline fallback for the canonical text-only positions (T=H=W=arange). |
| 5 | patch_cudagraph_align.py |
Aligns spec-decode CUDA graph capture sizes for pure PIECEWISE mode. Default FULL_AND_PIECEWISE spec-decode deployments already capture FULL decode graphs and have not reproduced this failure in long soaks. |
| 6 | ENV VLLM_TEST_FORCE_FP8_MARLIN=1 |
v1/v1.2 compatibility guard only. Current v2 images set this to 0 and use FlashInfer CUTLASS NVFP4 successfully on GB10. Keep Marlin only for older bases or shapes that still reject CUTLASS/grouped kernels. |
| 7 | ENV TORCH_CUDA_ARCH_LIST="12.1a" |
Native sm_121a (GB10) build target — the v0.23.0 image compiles 12.1a SASS directly (older image used sm_120+PTX JIT). |
| 8 | flashinfer 0.6.8 | sm_120 NVFP4 KV-cache decode kernels (PRs #2520, #2702). |
All patches live in patches/ and run automatically at image build time (idempotent). The Dockerfile is reproducible — see docs/build.md.
Note (current): the per-revision
vllm-spark-omni-q36lineage (v1/v1.2/v2) is superseded by the unifiedghcr.io/aeon-7/aeon-vllm-ultimate:latestimage for all production use. Pull that image — it bakes the GB10 CUTLASS NVFP4 defaults and the v0.23.0 fix set described below.
All AEON models run on one unified container — ghcr.io/aeon-7/aeon-vllm-ultimate:latest (= :2026-06-18-v0.23.0-dflashfix; rollback :2026-06-11-pr41703) — vLLM v0.23.0 built from source for GB10 / sm_121a and merged with the AEON speculative-decoding stack.
| Fix | What it does | Why it matters on GB10 |
|---|---|---|
| DFlash high-concurrency fix (new) | Slices the speculative drafter's KV block-table to the unpadded batch (block_table[:num_reqs]) |
The drafter previously crashed at ≥32 concurrent requests (padded-vs-unpadded block-table shape mismatch in FlashAttention). Now scales cleanly to c=64. A port of upstream PR #43982 (fixed for MTP, never applied to DFlash) — present and unfixed even in the prior image. |
| Triton NVFP4 KV cache (PR #44389) | Software NVFP4 KV-cache path | The only 4-bit KV path on sm_121a (upstream's is hard-gated to B200) → ~3× KV capacity / longer context per GB of unified memory. |
| DFlash sliding-window attention (PR #40898) | Runs a drafter's SWA layers as true sliding-window | Ships in the unified image for SWA-based drafters. This model's drafter is all-full-attention so it doesn't depend on it, but long-context draft acceptance holds regardless (42–58% at 32k, see above). |
| Prefix-cache-safe DFlash (PR #41703) | Makes --enable-prefix-caching corruption-immune under DFlash |
Lets the recommended recipe keep prefix caching on without garbled output, important for repeated long agent histories. |
| sm_121a-native build | TORCH_CUDA_ARCH_LIST=12.1a, ENABLE_NVFP4_SM100=0 |
Compiles the SM120-family CUTLASS NVFP4/FP8 kernels GB10 actually dispatches to — true 4-bit tensor-core throughput, no dead B200-only kernels. |
| sm_121a boot + CUDA-graph patches | RTLD-lazy _C_stable_libtorch load; spec-decode CUDA-graph capture-size alignment |
Boots past MXFP4 (SM100-only) symbols absent on GB10; prevents cudaErrorIllegalAddress on partial-acceptance decode steps under speculative decoding. |
| Unified-memory tuning | --gpu-memory-utilization ≤0.70, FULL CUDA graphs, async scheduling, z-lab DFlash drafter |
GB10 shares one LPDDR5X pool across CPU + GPU; conservative KV headroom avoids page-thrash while keeping FULL-graph + speculative-decode throughput. |
The result for this model: scales to 64 concurrent requests with no crash (the prior vllm-spark-omni-q36:v1.2 image crashed at c≥32 under speculative decoding), holds DFlash draft acceptance from short prompts through 32k-token agent histories, and serves at ~75–132 tok/s single-stream / ~700–800 tok/s peak aggregate (acceptance-driven across categories), and — as of the 2026-06-18 vision fix — multimodal image input works (validated 7/7). A stock vanilla-vLLM contrast for this model is still pending a fresh fully-vanilla re-benchmark, so no stock-vs-optimized speedup multiple is quoted yet.
Previous v1 weights had language_model. prefix stripped from safetensors keys to match a text-only model class — required vLLM registry + key-rename patches and was unstable in production (intermittent cudaErrorIllegalAddress crashes during real chat sessions).
v2 (current) re-quantized from tvall43/Qwen3.6-35B-A3B-heretic directly with AutoModelForImageTextToText, preserving:
- Full multimodal architecture (
Qwen3_5MoeForConditionalGeneration) - 27-block ViT vision encoder (BF16, NVFP4-skipped) —
⚠️ these vision tensors were initially mis-nested undermodel.language_model.visual.*and silently skip-loaded by vLLM (image inputs →!!!!) until the 2026-06-18 vision fix (below) - Original
model.language_model.layers.X.*key layout — vLLM's multimodal class loads natively, no prefix-strip patch needed - 30 linear_attention (Mamba/GDN, fp32) + 10 full_attention layers
- 256 routed experts × 8 active + 1 shared expert per layer
- All 122,880 per-expert NVFP4 keys (every expert calibrated)
vLLM serves it via the canonical multimodal class — fewer code paths in the inference hot loop, much better stability under load. Travis ran multiple live chat sessions (Celina) without a single crash where v1 was crashing on virtually every interaction.
v2 preserved the ViT's BF16 weights, but a quantization-time ignore-regex nested the 333 vision tensors one level too deep — as model.language_model.visual.* (a child of the language model) instead of the sibling model.visual.* that vLLM's Qwen3_5MoeForConditionalGeneration expects. vLLM silently skip-loaded the entire vision tower and ran it uninitialized, so text was perfect but any image input returned !!!! garbage.
The fix is a header-only rename of those 333 tensors to model.visual.* — the NVFP4/BF16 weight data is byte-for-byte identical (no re-quantization). vLLM now loads the full ViT (0 skip-loads) and image understanding works, validated on aeon-vllm-ultimate:latest (7/7 on a shapes/colors probe + a coherent scene description). The corrected model.safetensors is live on the HF repo — re-pull if you cloned before 2026-06-18. (The 27B siblings already used the correct model.visual.* layout and were unaffected.)
The compose serves 3 model aliases for the same backend:
qwen36-35b-heretic— canonical nameqwen36-fast— intended for greedy/agentic workloads (T=0 → high DFlash acceptance; 120+ tok/s single-stream on math/code/reasoning prompts)qwen36-deep— intended for sampled/creative workloads (T=0.7 → lower DFlash acceptance, ~75 tok/s on open-ended prose; tradeoff for diversity)
OpenClaw config (validated against actual zod schemas) is in docs/openclaw.md. The pattern: register two model entries pointing to the same backend with different default params.temperature. Route per agent or per channel binding.
| Doc | Audience |
|---|---|
docs/dgx-spark-setup.md |
Primary deployment guide — start here for full Spark setup |
docs/openclaw.md |
OpenClaw gateway integration (validated against real zod schemas) |
docs/dflash.md |
DFlash speculative decoding tuning + monitoring |
docs/dtree.md |
Future-work — slot DTree in when z-lab releases |
docs/quantization.md |
Recreating the NVFP4 quantization end-to-end (including v2 recipe) |
docs/build.md |
Building the image from source (advanced — canonical Dockerfile/patches live in the container repo) |
docs/troubleshooting.md |
Symptoms → root causes → fixes |
docs/patches.md |
Each patch explained, with the upstream issues they address |
- vLLM — vllm-project/vllm
- DFlash — z-lab/dflash (Soroush Mohri et al.)
- Qwen3.6-35B-A3B-heretic base — tvall43/Qwen3.6-35B-A3B-heretic (
heretic v1.2.0abliteration of unsloth/Qwen3.6-35B-A3B) - Qwen3.6 — Qwen team
- llmcompressor — vllm-project/llm-compressor
- FlashInfer — flashinfer-ai/flashinfer
- rmagur1203/vllm-dgx-spark — independent 4-day SM121 investigation that surfaced the Marlin requirement
- OpenClaw — openclaw/openclaw (Peter Steinberger / @steipete)
Apache 2.0 (matching upstream vLLM, FlashInfer, llmcompressor).
The base model carries its own license — see tvall43/Qwen3.6-35B-A3B-heretic.
If this release has been useful, tips are deeply appreciated — they go directly toward more compute, more models, and more open releases.
₿ Bitcoin (BTC) bc1q09xmzn00q4z3c5raene0f3pzn9d9pvawfm0py4
|
Ξ Ethereum (ETH) 0x1512667F6D61454ad531d2E45C0a5d1fd82D0500
|
◎ Solana (SOL) DgQsjHdAnT5PNLQTNpJdpLS3tYGpVcsHQCkpoiAKsw8t
|
ⓜ Monero (XMR) 836XrSKw4R76vNi3QPJ5Fa9ugcyvE2cWmKSPv3AhpTNNKvqP8v5ba9JRL4Vh7UnFNjDz3E2GXZDVVenu3rkZaNdUFhjAvgd
|
Ethereum L2s (Base, Arbitrum, Optimism, Polygon, etc.) and EVM-compatible tokens can be sent to the same Ethereum address.