Treasury funded from ETH → stock tokens
Swaps into all eleven payout tokens, weighted toward NVDA, TSM, AAPL and SKHY, delivered straight to the contract. Growing with deposits.
Everything shipped, newest first. Planned work is on the roadmap; this page only lists what is live.
Swaps into all eleven payout tokens, weighted toward NVDA, TSM, AAPL and SKHY, delivered straight to the contract. Growing with deposits.
fabrwa.xyz/runs/{id} for every verified run: gpu, sustained score, duration, lineage split, epoch, never the wallet. The feed and the leaderboard link to it, the result screen copies the link, and the link unfurls into a card drawn from the record.
fabrwa.xyz/status: which models have a GPU serving now, queues, answers checked today, the next settlement. fabrwa.xyz/gpus: every GPU measured under a mining pass, median sustained score per instrument, trial runs excluded.
The tab keeps the screen awake for the length of a run and lets go when it ends. An opt-in checkbox on the preflight screen sends a browser notification with the score when the run finishes.
One button under every miner answer in chat, or POST /api/v1/inference/{job}/verify: the same prompt goes to a different GPU and the verdict line names it. Greedy decoding, so an honest replay is the same text token for token. Free, ten checks an hour.
fabrwa.xyz/docs: twelve sections written from the code, every number read from the constants the fab runs on.
Trial runs are scored but not paid. The server sets the payout split, a run pays the wallet that signed its pass, and that wallet must hold $FAB when the run starts and when it finishes. Terms of use and a privacy policy that match the code.
Four ways in: mine, build, hold, watch. The long explainer moved to fabrwa.xyz/overview.
The miner's tab posts what it has about once a second while it decodes, relayed to whoever asked as server-sent events. Decoding stays greedy, so a streamed answer is still reproducible token for token on a second GPU.
t.me/siliconfab for updates, t.me/fabrwachat to talk, and a support page that says plainly what the fab can and cannot see about you.
fabrwa.xyz/chat for people, and /api/v1/chat/completions for anything that already speaks OpenAI. Multi-turn conversations, and every answer names the GPU or the provider that produced it.
DeepSeek-R1-Distill-Qwen-1.5B, 3.5 GB of weights from Hugging Face onto a laptop GPU. It needed q/k/v biases, an untied output head and a different pre-tokenizer, then we checked it against a numpy implementation written from the config file: same token ids, same text.
A node registers by signing with a wallet holding $FAB, declares which models it can actually serve, and answers from the same queue as browser miners. Qwen3-8B is the first rung. Nodes are not paid yet and not cross-checked.
Claude Opus 5 and DeepSeek V4-Pro answer through their maker's API on the caller's own key, used once and never stored. No markup, never counted as miner work, and the fab holds no provider key of its own, so none of it can bill the operator.
SmolLM2 at 135M, 360M and 1.7B behind one set of dimension-templated WGSL kernels, at full 16-bit precision. Weights stream into the tab in 192 MB range requests instead of one long download.
One in four answers is replayed on a second GPU and compared; the result is stamped on the response. $1 of ETH buys 1,000 prompts with longer answers, and the ETH goes to the treasury.
Hold $300 of $FAB when you take your pass and every run pays ×1.25. Hold $3,000 and it pays ×1.5. The tier is in the pass, checked on chain.
fabrwa.xyz/api/dataset: every verified run as CSV or JSON lines. GPU, vendor, instrument, sustained score, duration, country, epoch. No wallets, no IPs.
Mining pass for wallets holding $30 of $FAB, one free trial per ip, sustained scoring over long runs.
A path tracer and a transformer forward pass, verified and paid like every other instrument.
A buyer pays ETH and names a scene. Miners on FAB-03 render it while they mine, a frame every 20 seconds lands on the jobs board, the ETH becomes stock tokens in the treasury.
SmolLM2-135M-Instruct, bf16 weights fetched straight from Hugging Face into the miner's browser, 32 tokens per second on an M2 Pro. Greedy decoding, so a second GPU can recheck an answer.
POST a prompt to /api/v1/inference, a miner on FAB-05 answers it on their own silicon. Free tier, 20 prompts an hour, 256 tokens max.
Holographic bee mark, glass Fab banner, ink-violet ground with drifting blooms, brand colors per ticker.
Signed run tokens, per-vendor ceilings, rate limits, Upstash-backed global feed and per-model leaderboards.
Streaming bandwidth and tiled SGEMM instruments, GPU timestamp timing so browser jank cannot move a score.
0x2232…a68f, source verified. deposit(), closeEpoch(), claim() with settler signatures. First epoch closed on chain the same night.
Dollars owed become shares at settlement prices, vouchers signed per wallet, rate follows the vault: payouts are steered to a fixed share of holdings a day.
1B supply on the bonding curve, liquidity locks at graduation. $FAB is a key, not a reward.
WebGPU FMA workload, the wafer that fills in stepper order, lineage detection, runs, treasury, about.