Fleet List

Updated 25 October 2026

This page is the Fleet List and the list of available models that the Hourly Terms (clauses 1.1 and 8.2) and the 2-Year Reserved Fleet Service Agreement (definitions and clause 10.1) refer to. Each fleet's rate is the number of hours one hour on it consumes. The Korean version (fleets-ko.html) has the same content and prevails if they differ.

Fleets

Our system publishes live capacity, with each fleet's state and the hours open in each window, at https://sparkpipe.ai/availability.txt, regenerated at least every 30 seconds.

Models on Fleet 1

Every model installed on Fleet 1. An offered model can be chosen for any reservation, and named among the Chosen Models of a two-year agreement. A core model gets ongoing performance work; a qualified model passed the release qualification and serves at the speeds shown; a verifying model serves and passes the chat checks while its tokens are compared with the publisher's reference implementation.

Context windows are those served. Decode is one request without speculation, measured on 2026-10-18 with prompts of 7 to 8K tokens. Methods and receipts: PERFORMANCE_STATUS.md.

¹ DeepSeek V4.1 Flash was measured on 2026-10-24 with a prompt of 1,047,883 tokens: decode one request after the prompt, and the prompt's average prefill rate. ² Measured while another workload kept 16 CPU cores busy on every node, so these are lower bounds. ³ Measured on 2026-10-26 with prompts of 6,850 to 7,396 tokens. ⁴ With speculation drafting from the request's own text, a file of about 3,800 tokens rewritten whole with one function renamed runs at 190.9 tok/s on K-EXAONE 2.0, 298.0 on A.X K2 and 426.6 on Solar Open 2, prefill included, with output identical to decoding without speculation (2026-10-26).

Qwen 3.8 Flash is not offered: its license requires a separate agreement with Qwen for any model-as-a-service business, so we use it only inside our company.

Speculation (measured)

Speculation drafts tokens from the request's own context and from a draft model, and the model checks them in one pass, so the output is exactly the model's own. Measured on Fleet 1, one request unless noted:

Receipts: PERFORMANCE_STATUS.md, 2026-10-15 to 2026-10-24.

Near-term speeds for the largest and Korean models (ESTIMATED)

ESTIMATED values are projections from our measurements and the hardware's published figures, not measurements. We publish one only when we expect to meet it, and each is replaced by a measurement when the hardware or software behind it is running. Cells marked measured already are.

Measured on one M5 Ultra Mac Studio, the Fleet 2 machine, with SparkPipe's Metal engine (2026-10-26): Qwen 3.8 27B at its full BF16 precision decodes 20.1 tok/s for one request, 90% of the 22.3 tok/s that the Studio's measured 1.14 TB/s of memory bandwidth allows. Several requests share each read of the weights: 39.0 tok/s in total for 2 requests, 130 for 8, 231 for 16, 366 for 32 and 442 for 64. Its prefill runs at 1,429 tok/s for a 1K-token prompt and 1,501 tok/s for 8K.

A 1M-token Kimi K3 prompt from cold is estimated at about 17 minutes with four accelerator boxes and 9 minutes with eight. Fleet 1 and Fleet 2 together hold 4 TB, and Fleet 3 adds 2 TB more, so a model too large for one fleet at its published precision can run across two: Qwen 3.8 Max in FP8 (2.5 TB) and Inkling in BF16 (1.9 TB).

How the estimates are made:

• Decode: a token takes the time each node needs to read its share of the weights the token uses, at the 245 GB/s measured on Fleet 1, plus a fixed 0.44 ms per layer for the kernels and collectives around those reads. Kimi K3 measures 0.44 ms and GLM 5.3 0.31 ms; that fixed cost, not memory, is most of today's decode time, so the estimates rise as our work lowers it. Fleet 2 is taken at 90% of the Mac Studios' 1.2 TB/s with the same cost per layer and the model split over all eight, which are joined by 50 Gb/s links plus 10 GbE. Qwen 3.8 Max's 10.4 tok/s is not yet explained by this, so it has no Fleet 2 estimate. K-EXAONE 2.0's Fleet 2 estimate is being redone now that its experts are split evenly over the nodes.
• Speculation on agent coding turns is taken as 1.5 times plain decode, as measured on short edits; whole-file rewrites measure 8.3 times and new prose 1.0 to 1.2 times.
• Prefill accelerators are boxes of four RTX 5080 cards that reduce over PCIe inside the box and reach the fleets at 100 Gb/s each over a 400 Gb/s switch. Each long prompt streams the model's weights through the boxes once, at 85% of the links (42.5 GB/s with four boxes, 85 GB/s with eight), and the cards compute at 40% of their dense peak (about 450 TFLOPS each in FP8, 225 in BF16), including the attention cost of the prompt. Prompts shorter than about 20,000 tokens stay on the fleet's own prefill.
• Qwen 3.8 Max at its published FP8 precision (2.5 TB) is larger than one fleet's 2 TB. Its row assumes one of its 4-bit builds (AMD MXFP4, Red Hat or RadixArk NVFP4), which fit one fleet and pass our accuracy A/B.

Models coming next

Large open models from non-Chinese publishers, under licenses that allow serving them to businesses. We qualify each one before it is offered.

Model licenses

Each model is the publisher's open-weight software under its own license, which the customer must follow (Hourly Terms 8.1, agreement 10.3). The conditions of the offered models:

• Kimi K3: a model-as-a-service business whose group revenue exceeds US$20 million in any 12 months needs a separate agreement with Moonshot AI, and a commercial product with more than 100 million monthly active users or US$20 million in monthly revenue must display "Kimi K3" in its interface.
• GLM 5.3: a model-as-a-service business whose group revenue exceeds US$10 billion in any 12 months must pass Z.AI's security review before commercial use.
• Qwen 3.8 Max: a model-as-a-service or AI work-assistant business whose group revenue exceeds US$50 million in any 12 months needs a separate license from Qwen, and a commercial product with more than 100 million monthly active users or US$20 million in monthly revenue must display the model's name.
• MIT (GLM 5.3 Flash, DeepSeek V4.1 Flash, MiMo V2.6, Ling 3.0) and Apache 2.0 (Gemma 4, Muse Glimmer 30B, Hunyuan Hy4 Preview, Qwen 3.8 27B, K-EXAONE 2.0, A.X K2): no field-of-use or revenue condition.
• Upstage Solar License (Solar Open 2): the Apache 2.0 terms, plus a rule that a model derived from Solar Open 2 and distributed to others must state the "Solar" brand; no field-of-use or revenue condition.
• OpenMDW 1.1 (Laguna S 2.1): no condition on use or outputs; the license ends for anyone who sues claiming the model infringes a patent or copyright.

The Kimi K3, GLM-5.3 and Qwen 3.8 Max licenses also require every use to comply with applicable law. Every license requires the publisher's notices to stay with any copy of the weights; customers use the models on our fleets and receive no copy.