Total seats
70
18 dedicated entitlements + 52 free assignments
TangTrades research · 2 November 2026
Distributed AI.
Different models.
MLX alongside Apple Minions, EXO, Darkbloom, Pearl Research and Bittensor. Architecture, commercial evidence and token economics.
Public network telemetry
Seats, Macs and work.
Source time: 2 Nov 2026 · 02:44 SGT
Macs seen in 7 days
64
Recent linked devices, including those now offline
Laptops seen in 7 days
45
MacBook entries; 19 desktop Mac entries separately
Macs online
15
14 working or waiting between jobs
480Reported tokens / second
177.22MRecorded tokens · 7 UTC-day window
0Paid API requests · same window
252.3 TFLOPSPledged class estimate · not a benchmark
Saved snapshot; checking the public APIs for a newer observation.Work counter: 2 Nov 2026 · 02:44 SGTLedger: 2 Nov · 02:35 SGT
Seats are entitlements and assignments; they are not a computer count. Recent devices are project-reported entries, not an audited inventory of unique hardware. Live counters refresh while this page is open; charts and research commentary retain the dated snapshot. Metric definitions ↗
What changed since the previous published activity snapshot?
Comparison with 1 November, 19:12 SGT, using the new 2 November, 02:44 SGT research snapshot. These are dated observations; the live cards above can be newer.
| Measure | Previous | Updated | Change |
|---|---|---|---|
| Total seats | 72 | 70 | -2 |
| Macs seen in 7 days | 67 | 64 | -3 |
| Linked registration keys | 81 | 81 | Unchanged |
| Recorded work tokens | 164.65M | 177.22M | +7.64% |
| Mining distributions | 1,080.80 MLX | 1,110.61 MLX | +29.81 MLX |
| Paid API requests | 0 | 0 | No paid requests recorded |
A rolling-week count can decline when older entries leave the window. This does not identify why individual devices stopped reporting. The increase from 305.4 to 309.9 estimated registered TFLOPS is a parsing correction for an existing Apple M5 Pro entry, not newly added hardware.
03 / COMPETITIVE LANDSCAPEProduct architecture · commercial considerations
Six approaches to distributed AI.
01
MLX.app
Mac software + token network
Product and approach
MLX Node contributes AI work; Harness runs local and connected AI tools. The project publishes token and supply data. [4][1]
Key considerations
Commercial scale depends on repeat paid workloads, reliable execution and sustainable worker economics. Published reserve figures remain project-reported; independent reserve verification has not been established. [5]
Product and approach
Apple research pools idle Macs for independent embedding and text-generation tasks. Workers claim jobs through a central service; interrupted work can be retried. [12][13]
Key considerations
Reports shorter internal GPU queues. The paper describes no token or public compute marketplace. A technical comparable; public-network economics and customer demand require separate evidence. [13]
03
EXO
Local model clustering
Product and approach
Open-source software discovers devices and splits models according to available resources and network topology. [6]
Key considerations
Fast multi-Mac configurations rely on compatible hardware and local networking. It is a reference for local clusters; geographically distributed devices face different latency and bandwidth constraints. [6]
04
Darkbloom
Routed inference on idle Macs
Product and approach
Routes API inference to Apple Silicon providers, with per-token pricing and published trust boundaries. [7]
05
Pearl Research
Inference serving + mining incentives
Product and approach
Offers serverless, managed and on-premise inference. Together announced a Pearl-powered endpoint using token incentives to subsidize inference pricing. [9][10]
Key considerations
Token emissions can subsidize delivery costs. Durable commercial economics still depend on paying customers, service reliability and margins after incentives.
06
Bittensor
Subnet incentive network
Product and approach
Independent subnets produce digital services; miners supply outputs and validators score them within a token-incentive framework. [11]
Key considerations
Broader than a single inference provider. Customer adoption, service quality and economics need to be evaluated at the subnet level.
04 / TOKEN COMPARISONUSD · 2 Nov 2026 · MLX read 03:04 SGT · Market API 03:02:58 SGT
Price, supply & liquidity.
A consistent maximum-supply FDV calculation, with different pricing mechanisms made explicit. These are dated market observations; live refresh above applies to network telemetry.
MLX
MLX ↗Protocol curve · issuer-reported
- Circulating market cap
- $20,481
- FDV · maximum supply
- $100.1M
- Reference price
- $0.01001
- Circulating supply
- ≈2.0M MLX
- Maximum supply
- 10B MLX
≈0.020% circulatingof maximum supply
Curve-implied FDV. The issuer reports $23,788 of USDC sell liquidity; reserve verification remains unestablished. The reference curve price is not the executable sell quote.
Pearl Research
PEARL / PRL ↗External-market quote · CoinGecko USD API
- Circulating market cap
- $554.3M
- FDV · maximum supply
- $3.192B
- Reference price
- $1.52
- Circulating supply
- 365.355M PEARL
- Maximum supply
- 2.1B PEARL
17.40% circulatingof maximum supply
FDV = $1.52 × 2.1B tokens. Token incentives and commercial demand require separate evaluation.
Bittensor
TAO ↗External-market quote · CoinGecko USD API
- Circulating market cap
- $2.716B
- FDV · maximum supply
- $4.879B
- Reference price
- $232.31
- Circulating supply
- 11.682M TAO
- Maximum supply
- 21M TAO
55.63% circulatingof maximum supply
FDV = $232.31 × 21M tokens. Maximum-supply FDV differs from the aggregator’s issued-supply FDV.
Valuation basis. Circulating market caps are source-reported. FDV = displayed reference price × maximum supply, so it can differ slightly from a provider’s figure using an unrounded price. Supply and market-cap fields can update independently. MLX’s circulating supply is rounded by its source; the comparison does not infer a more precise float from it.
Price and liquidity. MLX’s protocol curve is distinct from external-market prices for PEARL and TAO. Its published buy premium is 3% and sell discount is 20%; the displayed spot buy / sell quotes are $0.01031 / $0.008007. Neither circulating capitalization nor FDV is the cash available for exit. Public reserve amounts are issuer-reported.
Sources: MLX token page, explicit USD market API, captured market data ↓. The source manifest preserves the previous comparison snapshot as historical provenance.
05 / APPLE MINIONS × MLX.APPPrimary sources rechecked 2 November 2026 · Singapore time
The same spare Macs.
A different operating model.
Apple Minions turns spare Mac GPU capacity into a shared inference service. MLX.app combines outside participants, wallets, rewards and customer billing. [12][14][15]
WHAT APPLE BUILT
Published in October 2026, Minions divides AI work into independent batches that available Macs execute. Its described workloads are embeddings and LLM completions, using Hugging Face transformers and Ollama. [12][13]
This is Apple's Minions paper, separate from Apple's MLX framework and the independent MLX.app network. MLX.app states that it has no affiliation with Apple. [12][19]
| Dimension | Apple Minions [13] | MLX.app [14][15][16] |
|---|---|---|
| Primary purpose | Clear teams' inference backlogs using existing Macs. | Recruit Mac providers and develop a paid compute service alongside token mining. |
| Coordination | Workers pull tasks through a central server; PostgreSQL tracks ownership; S3 holds data. | Distributed workers receive jobs from a central server that collects results and applies reward rules. |
| Work available | Embeddings and LLM completions. Distributed training is future work. | Published paid pilot: Qwen 2.5 0.5B, text only, up to 120 output tokens, no streaming. [18] |
| Trust & privacy | Organization or team pools can be isolated for confidentiality. | Raw customer prompts go to other participants' Macs. The pilot advises nonconfidential work. |
| Failures & verification | Heartbeats and a watchdog return unfinished work to the queue. | Normally compares two executions from different wallets in the same GPU group; may use a third after disagreement. After 45 seconds, a lone response can be billed as unverified. Matching output measures consistency, not independent answer quality. |
| Payments | No token or customer-payment design described. | Paid jobs go only to dedicated-seat Macs. Customer fees: 90% to matching workers in USDC; 10% to treasury. Token mining is separate. |
Coordination matters. Minions avoids a central push scheduler by letting workers claim tasks, but retains central services. MLX also has a central job server. Calling one centralized and the other fully decentralized would obscure the documented architectures. [13][16]
APPLE · REPORTED OPERATIONAL RESULT
Hours of queueing → minutes
Some jobs previously waited over ten hours; Minions often let them start within minutes. This measures access to compute, not faster model execution. The paper supplies no fleet count, throughput or full operating-cost benchmark. [13]
Reusing hardware avoids new GPU purchases. Electricity, storage, networking and operations still need to be costed.
MLX · PUBLIC COUNTER SNAPSHOT
Paid usage remains to be demonstrated
The public counter reported zero paid API requests and zero paid tokens in its displayed seven-day window. It described recorded token work as verification activity. [17]
Research snapshot: 2 Nov 2026, 02:44 SGT. The counter reported 177.22M tokens, all verification work. Project-reported and unaudited; the live panel at the top can show newer observations.
IMPLICATION FOR MLX · ANALYTICAL JUDGMENT
Technical feasibility is supported. Commercial differentiation is still the task.
Minions strengthens the case for useful work on spare Macs. MLX's defensible value would come from distribution, repeat customers, reliable delivery, privacy, verification and sustainable provider earnings. Connecting Macs alone offers limited differentiation.
We would prioritize nonconfidential, retryable batch work such as public-document processing and embeddings. Embeddings are a proposed MLX expansion here; the current paid pilot documents chat completions. Neither cited system demonstrates one frontier model split across a worldwide Mac fleet.
Confidence: high on the documented architectures and published limitations; moderate on this commercial interpretation. Apple Minions is included as a technical comparable. No standalone token valuation is assigned.
Current product / published scope
The pilot, precisely.
The local Harness app and the paid network API have different capabilities. This panel describes the published paid API.
| Model & price | Qwen2.5-0.5B-Instruct, 4-bit · $0.02 per million prompt or completion tokens |
|---|---|
| Request limits | 4,000 prompt characters; up to 120 output tokens; no streaming; up to 45 seconds |
| Routing | Dedicated-seat Macs. Full eligibility depends on wallet, GPU group, availability and worker state. |
| Verification | Same-group output agreement; a third Mac can resolve disagreement. A lone response can be billed as unverified at the deadline. Execution audits are partial and do not currently change pay. |
| Customer payments | 90% to qualifying workers and 10% to treasury in USDC. MLX mining distributions are a separate flow. |
| Privacy | Workers receive customer prompts. Initial customer tests should use nonconfidential data. |
| Observed commercial status | Pilot open since 27 October. At the 2 November research snapshot, the public seven-day counter shows zero paid requests and zero customer spending. |
Sources: model catalogue, request limits, routing and payments, execution checks. Published specifications were checked; they do not constitute an independent service benchmark.
Evidence / methods / downloads
Make the numbers checkable.
The public APIs provide the observations. The downloaded datasets preserve coverage, source timestamps and calculation methods.
World network snapshotRecent Macs, online activity, seats and token rate↗Work & paid-usage countersExecution tokens, paid requests and customer spending↗Activity dataset345 hourly records, reward reconciliation and daily tokens↓Network datasetDevice events, seat records and 1,892 minute samples↓Ledger & seat statisticsDedicated entitlements, chain checkpoint and subsidies↗Public metric definitionsWindow length, activity states and measurement limits↗
Research cutoff: 2 November 2026, 02:44 SGT. Reward ledger checkpoint: block 788. The 345 hourly records reconcile to 1,110,613,036 MLX base units, or 1,110.613036 MLX. The prior 338 records are retained unchanged.
Chart coverage: registered keys and dedicated seats are reconstructed from the ledger. Token-output counters start 27 October. The minute archive retains 1,892 observations; no earlier online history is fabricated. Displayed daily token dates are UTC, while registration and hourly-reward dates use Singapore time.
Refresh behavior: headline metrics query the public APIs while the page is open. If retrieval fails, the last successful values remain visible with their source times. This does not collect background history; the downloadable datasets and research commentary remain at the dated cutoff. External token-market observations have their own source timestamps.
Source context: Network and reserve figures are project-reported. MLX.app is independent of Apple; Apple Minions and Apple's MLX framework are separate projects.
Comparison references and primary documentation
- MLX token: price and supply
- Pearl Research market data
- Bittensor market data
- MLX product and applications
- MLX reserve disclosures
- EXO source code and setup
- Darkbloom architecture and trust boundaries
- Darkbloom documentation and development status
- Pearl Research inference services
- Together AI and Pearl Research partnership
- Bittensor network documentation
- Apple Minions: official research publication
- Minions paper: architecture, operational results and limitations
- MLX Compute API: current pilot scope
- MLX Compute API: routing, verification, payment and privacy
- MLX job protocol: central coordination and verification limits
- MLX public compute-work counter
- MLX Compute API: model and request limits
- MLX Network and Apple MLX: affiliation and runtime