Eleanor is our resident researcher. Click the planet to speak with her.

Frontier-class AI you can own.

Brainsless is a research lab that makes owning a frontier-class model practical. We take the strongest open models and make them ours: our own training, our own memory system, our own serving stack, specialized to a job and run on hardware we control.

Today that stack serves a trillion-parameter model at over 510 tokens per second from four GPUs. That is above every provider median the public leaderboard publishes for this model, at the leaderboard's own workload, and it is measured, published, and reproducible down to the prompt seeds.

A faster step is not only lower latency. It is more tokens from the same GPUs, so one cluster carries more users before it saturates. That is the lever that turns owning a frontier model from a hyperscaler's budget line into something a focused team can run, and it is the difference between renting intelligence and holding it.

The lab exists to break the walls that make ownership impractical: serving speed, serving cost, and specialization. Planless, our AI co-founder, is where the proof runs. It is trained on the work of building a company, it remembers through Cortex, our memory system, and it answers from our own model on our own stack.

Four technical reports back the research. One lab, one thesis: a specialized model you own beats a general one you rent.

Own the stack

Model, memory, drafter, serving engine: every layer is ours to measure and ours to change. When a number moves, we can see which layer moved it.

Measure, then claim

Every number we publish rebuilds from a named artifact: prompts, seeds, serve commands, raw outputs. Claims ship with the code and data to re-run them.

Prove it in production

Planless runs on this stack, under real use and real stakes. The workloads we benchmark are the workloads it serves.

What we build
Fovea
Train a draft head on your own traffic and serve your model faster, with identical outputs. In active development, serving our own production today.
See Fovea · join the waitlist →
Cortex
The memory behind the co-founder. It holds who you are the way a person does, not as a transcript search.
How it works →
Planless
An AI co-founder that remembers your company and carries the work with you. Where the research reaches real people.
planless.app ↗
Neptyn
Our production model. An open-weight foundation we extended with our own training, memory, and inference systems.
See the model card ↓

The research

BRL-2026-11 · NEWJuly 2026

Over 510 tokens per second on Kimi-K2.6, from four GPUs.

511.6 tokens per second, single stream, lossless — a record set at 505.9 (the first GPU reading above 500 on this model) and raised by blind re-runs of the unmodified public release. Replication moved this record up, not down. Above every provider median the leaderboard publishes for this model (fourteen providers listed) (best at the record's pin: 438.1, from undisclosed hardware; 451.0 at the raised cell's pin, with 511.6 standing 13.4% above). Per GPU about three times NVIDIA's disclosed trillion-parameter configuration. The report carries the decomposition that found it: the forward pass costs 6.6 milliseconds, verifying a guessed token costs 0.4, and most of the rest of the step was engine overhead we removed. Tool-calling traffic sets the headline; the math cells read 419.2–422.1, and every session and node draw is printed. Verification is a purchasable fact: one command, about $15, re-runs the record from the public release. Every part in the configuration is stock and public: the number is a floor, and our trained drafter is already 5.8% ahead of the stock head in a same-session pair.

DOI 10.5281/zenodo.21217452
Read the record → The paper PDF Run it yourself
GPU serving of Kimi-K2.6, single stream, 10k-token workload — pinned 2026-07-05, n=16
External bars: Artificial Analysis, with NVIDIA's published 671B single-user record added for reference. No other entry runs a configuration as small as four GPUs.
The attention series — how little of a million-token context models actually read
First page of Attention Has A Type
BRL-2026-06June 2026

Attention Has A Type

The law. To predict its exact next word, a model only ever needs a small, fixed number of memories, whether the conversation is eight thousand tokens or a million.

DOI 10.5281/zenodo.20582700
Read → PDF
First page of Attention Pays Its Bill
BRL-2026-07June 2026

Attention Pays Its Bill

The cost paper. A long conversation now costs the same per word as a short one, and a million-token cache runs from about $1,500 of ordinary RAM instead of $180,000 of GPUs.

DOI 10.5281/zenodo.20673046
Read → PDF
First page of Attention Finds Its Keys
BRL-2026-08June 2026

Attention Finds Its Keys

The frontier. Past half a million tokens, models stop finding the right memory, and the rank law explains why: the collapse is arithmetic, not bad luck.

DOI 10.5281/zenodo.20673107
Read → PDF

Mohammad Alsufi & Connor Boone, with the Brainsless Research Lab AI Systems Research Group. Code public, noncommercial license. Re-run the record yourself →

The benchmark

GPU serving of Kimi-K2.6, single stream.

Our measurement against the field, pinned 2026-07-05. External numbers are Artificial Analysis live-endpoint medians at the 10k-token workload; ours is a server-side benchmark at the same workload, n=16 per cell. No other entry runs a configuration as small as four GPUs.

Deployment Hardware GPU count stated tok/s, single stream Measurement
Brainsless (ours) 4×B200, one node Yes — 4 511.6 Server-side benchmark, n=16, tool cell; record set 505.9 (2026-07-05), raised by blind re-runs of the public release; math 419.2–422.1
Crusoe Undisclosed No 438.1 AA live-endpoint median (peak 449)
Fireworks Undisclosed No 381.2 AA live-endpoint median
CoreWeave GB300 NVL72 rack (their blog) Rack-scale — 72 261.8 AA live-endpoint median
Nebius Undisclosed No 222.4 AA live-endpoint median
Together (FP4) Undisclosed No 218.2 AA live-endpoint median
GMI (FP8) Undisclosed No 40.2 AA live-endpoint median

Two reference points outside the table's scope. Cerebras serves this model at 981 tok/s from wafer-scale hardware on a private endpoint; it is excluded from a GPU comparison by construction. The best documented lossless single-stream number in this model class is NVIDIA's 340 tok/s on DeepSeek-R1 671B from eight B200s under TRT-LLM; their 368 variant relaxes acceptance and pays 2.8 points of MMLU-Pro, so it is a different claim. Ours is lossless: the full model verifies every token, and the output distribution is the model's own. On 2026-07-06 we replayed the same workload against one provider's production API from a standard paying account: it read 118.4–168.9 tok/s across its endpoint products against a 381.2 board median (fireworks_live_replication.json); the table above uses the board's numbers as published. The record itself replicated: three blind re-runs from the public release landed every n=16 cell inside the stated node envelope, raising the record to 511.6 (depth 6), with first-eight medians to 538 and single requests to 568 (brl11_repl_r1.json, brl11_repl_cleanroom.json, brl11_repl_fovea.json).

Protocol, ours: 16 distinct novel ~10k-token docpack prompts per domain, 2048-token outputs, temperature 0.6, per-request seeds, no prefix-cache reuse. Per-request streaming decode rate (ctok−1)/(t_last−t_first), interpolated median of n=16, Kimi-native tokens (o200k parity 1.0035–1.0081). Engine: SGLang v0.5.14, EAGLE3 speculative decoding (public 3B MLA draft head, draft depth 7, chain top-k 1), fp8 KV cache. Node draws move throughput ±5–10% and every session and draw is disclosed. Acceptance is cross-validated on two engines (τ 4.825 SGLang, 4.815 vLLM, equal depth, same head). External medians fetched 2026-07-05 from artificialanalysis.ai. Artifacts: brl11_stage600a.json, brl11_stage600e1.json, brl11_stage600b.json, brl11_stage600sg.json, brl11_stage600r.json (the record session).

Our model
Neptyn 1.0
The model behind the Planless co-founder
Foundation
Open-weight
Memory
Cortex
Drafter
Fovea
Routing
Instant + Thinking

Neptyn starts from an open-weight foundation and goes further with our own training, memory, and inference research. It remembers across sessions, carries its own state, and acts on your behalf.

Cortex gives it memory that consolidates like a brain. Fovea, a draft head we train on our own traffic, is part of why it answers fast. Two variants route automatically: Instant handles conversation, Thinking handles the hard calls.

Cortex → Fovea → Try Planless ↗