Every model skycmd publishes, one card each: what it is for, its file size, how it was judged, the one fact it brings and where to try it. Most were judged by a gate, a bar written down in a pre-registration before the model was trained; a model that misses its gate is published anyway, as a result. The numbers on the cards come from Results, What we bring and the Hugging Face model cards. Download links, formats and the engines each file runs in are on Install.
Each model also has a demo on Hugging Face, and all of them are gathered in one place: Every demo in one Space, where each demo opens with its model's verdict, passed or failed.
Passed
It cleared a gate written down before it was trained and tested.
Passed on a second look
It cleared its gate, but it was designed after a first version failed on the same test records.
No gate, measured
Its numbers are measured and published, but no gate was fixed for this file before it was tested.
Passed on its bench only
It cleared its gate on generated data and fails on real data, measured after publication.
Failed gate, usable as a part
It missed its gate; what it does measure makes it worth using inside another tool.
Failed gate, a result
It missed its gate. The file is published as the result of its experiment, not for use.
These models do what their card says, on the tests their card names. Four passed a pre-registered gate. Two, s3-pico and s5-lite, were measured without one, and their status says so.
s6-ear2-nano
632 KB
Passed
Detects the end of a speaker's turn for a voice agent: give it the audio, ask it once when the VAD reports a pause.
What it brings
The right answer on 0.8993 of the 31,527 test clips of Smart Turn v3.2 (bar 0.85; the teacher measured 0.9260 there, s6-ear-nano 0.5681, the silence timer 0.4979), at 2.4 ms per decision in WASM and 632 KB, where its teacher is 8.7 MB and about 32 ms natively.
Know before you use it
On real conversations (eot-bench) it and its teacher are far weaker than on the test clips, and at its threshold it answers "finished" on about half of the real ends of turn. It is weakest on Vietnamese, Marathi and Bengali. Its successor, trained on real conversations, failed its gate, so this stays the file to take.
Picks the tool that fits a request from a list of tools you write at run time, or says that none fits.
What it brings
The right option on 0.8868 of 945 fresh requests about tools never seen in training, against 0.8169 for BM25 and 0.9048 for all-MiniLM-L6-v2 (91.6 MB), in 30.9 ms per request in WASM.
Know before you use it
It needs distinctive tool names: with every tool renamed it falls to 0.2497. On fresh requests about MCP tools, a family unlike its training data, it read 0.5530 where BM25 read 0.5675 (reported with s6-route v7).
Status
Passed on fresh requests; the system was defined after its predecessor's test was read.
Plays four browser games (Flappy Drone, Rescue Run, Obstacle Dodge and a Doom-like) from the game state.
What it brings
The four games at the oracle's level (Flappy 200/200 pipes, Dodge 87%, the oracle's rate), at 49.6 µs per Flappy frame in WASM; trained from scratch in about 47 minutes on an Apple M5.
Status
Its run had no pre-registered gate; the scores are 30 seeded episodes per game with the shipped file.
Answers questions you write yourself, in English, with each option defined in plain words.
What it brings
Reordering the options or renaming their keys cannot change an answer, by construction (exact tests to 1e-6, natively and in WASM); 0.696 on the typed-decisions test, where oscar-1-17m scores 0.682 on the same records.
Know before you use it
26% of its answers flip when the definitions are paraphrased, and four test sets fall below their majority class.
Status
The pre-registered s5 bets failed; this checkpoint is a longer run added for the demo outside them, so its numbers are reported, not claimed.
It missed its gate, and what it does measure makes it worth using behind a rule-based scanner.
s6-secrets-verify-pico
250 KB
Failed gate, usable as a part
Judges whether a value found by a secret-scanning rule is a real credential, from the value, its name and the lines around it.
What it brings
Line F1 0.8561 on 221 real test repositories of CredData, above every rule scanner tested (the best, Betterleaks, 0.7332), with 0.35% of clean real files flagged; its gate asked for 0.8832.
Know before you use it
secretgate 1.9.0 offers it as an experimental option, off by default, where it can only add redactions. Its input comes from secretgate's rules, which the browser does not run: the demo proposes candidates with one simple rule instead.
Status
It missed one gate of seven (G1) by 0.027; the six others hold.
Models that failed their gate, or that a later model replaced. Each file stays published as the result of its experiment, and each one taught something a later experiment used.
s6-route-nano
2.4 MB
Failed gate, a result
Picks a tool from a list given at run time, with each tool's BM25 score in its input.
What it taught
It tied BM25, the score it reads (0.8395 against 0.8500 on 2,000 slates of tools never seen in training), at 38.9 ms per request in WASM where its gate asked for 2 ms. Its file, unchanged, is the model half of s6-route.
The tool router of s6-route with its model part retrained on data rebuilt against the cues that give the answer without the request, read on tools of a new family.
What it taught
On 2,000 fresh requests about the tools of MCP servers that no training slate holds, it reads 0.5610 where BM25 reads 0.5675 and its own validation read 0.8802. With the request hidden its trained part picks the right tool at chance (0.1406, chance 0.1429), and masking the names costs it 0.0170, where it costs s6-route 0.2720 on the same requests.
Know before you use it
A result, not a router to use: it does not beat BM25 on new tools, its ECE there is 0.1845, and the pair with the request hidden picks the right tool on 0.1969, above its bar of 0.1929. The line is closed; s6-route stays the router.
Status
It missed G1 (BM25), G2 (none under masked names), G3 (calibration) and G6 (b) (the pair with the request hidden); G4, G5 and G6 (a) hold.
An s6-nano with a fast-weight memory that carries a fact from one frame to a later one, and "sleeps" by replaying its last frames.
What it taught
A ternary model trained to use a fast-weight memory: right on 0.4892 of the recall test questions, against 0.2440 for the same training without a memory. The memory fades with the delay (0.6147 at 1 to 4 frames, 0.3302 at 17 to 32), four sleep passes add +0.0030, and it scores 15.87 at Flappy where the control scores 71.05.
Know before you use it
A research result, not a part to use: the memory fades with the delay, and the model plays Flappy and Dodge worse than the same training without a memory.
Status
Its memory is used (H0 holds); sleep passes did not help (H1) and two games lose (H2, and the s6-nano gate).
Says whether a short text exposes a secret, which service issued it and what kind of value it is.
What it taught
A generated bench does not carry over to real files: it beat a rule scanner there (secret 0.9990 against 0.977), then flagged 59.8% of 719 clean real files at p ≥ 0.9.
The first version of s6-micro: four games and typed questions in free text.
What it taught
Accuracy is not the whole gate: 0.547 on typed_decisions_bench cleared its bar of 0.45, and an ECE of 0.0821 missed the 0.08 guard. Its second version, with twice the decision-training minutes, passed on a second look.
s6-ear2-nano's end-of-turn graph trained again on real two-speaker conversations and on the silences a voice agent asks at.
What it taught
Real conversations taught it its own corpus, not speech addressed to an agent: on held-out human-to-human conversations its AUC is 0.6878 against v2's 0.6031, and on eot-bench, 0.2 s into each pause, 0.7197 against v2's 0.7448 (at least +0.02 was asked), where Smart Turn reads 0.7873.
Know before you use it
A result, not a detector to use: s6-ear2-nano stays the file to take. It was designed after v2's eot-bench reading, and its label cross-check was amended after a measurement on five training sessions.
Status
It missed E1, E2 and E3 (real speech addressed to an agent); A1, A2 and G3 (Smart Turn's clips, latency and size) hold.
Hears the last 10 s of audio and answers typed questions about it: is the turn over, which keyword was said.
What it taught
Its pretraining on Mimi speech codes did not help: a control trained without it did better on all four test sets. End of turn 0.5681, against 0.9263 for Smart Turn v3.2 on the same clips.
Forecasts the next 64 steps of a time series as nine deciles, with the periods it detects in the history as an extra input.
What it taught
A periodicity front end narrowed the gap and did not close it: nGMASE 0.8847 and nWQL 0.6190 on GIFT-Eval, against bars of 0.838 and 0.587, where s6-cast-pico scored 0.9471 and 0.6646 and our TinyCast port scores 0.7918 and 0.5622 without flip.
Know before you use it
A result, not a forecaster to use: our TinyCast port forecasts better on GIFT-Eval and on Chronos-ZS. GIFT-Eval is partly in-sample for its design choices.
Status
It missed G1, both GIFT-Eval bars; G2 holds (its MASE is under the seasonal naive at each term).
Forecasts the next 64 steps of a time series as nine deciles.
What it taught
nGMASE 0.9471 and nWQL 0.6646 on GIFT-Eval, against bars of 0.838 and 0.587; our port of TinyCast, a published model of 146,505 parameters, scores 0.7856 and 0.5570 on the same harness.
Generic decisions with options in free text, distilled from Qwen3.5-4B in one night.
What it taught
Loops do not replace language pretraining: 0.343 on the typed-decisions bench against a gate of 0.55, below the majority class (0.398); a fine-tuned 4.4M BERT-Tiny reaches 0.794 on Banking77, against 0.303.
Play Flappy Drone, Rescue Run and Obstacle Dodge; s2-nano-doom, its fine-tune, plays the Doom-like.
What it taught
Calibrate early exits by playing: at exit thresholds fitted on i.i.d. questions, s2-nano crashed in every Flappy episode; with thresholds chosen in the game it flew 191.9 of 200 pipes on average.
Plays three browser games in the standard System One format; the same file runs in llama.cpp and wllama.
What it taught
Labelling states is not playing: it labelled 82% of Flappy test states correctly and crashed in all 30 episodes; relative clearances in its input and two DAgger rounds took it to 200/200 pipes.
Pre-registered and not yet read on a test set: no score until the gate is read. Each pre-registration says what is tested, how, and what would count as a pass. No card is here: every line is closed (20 November 2026), and no pre-registration is waiting for its reading. The last ones to be read, s6-cast v2 on 18 November 2026, s6-route v7 and s6-ear v3 on 19 November, are now s6-cast2-pico, s6-route7-nano and s6-ear3-nano, under Published as results. s6-sleep v4 closed on 19 November at its check K1, on training sequences, with no model (docs/s6-sleep4-prereg.md §16); s6-sleep3-nano stays the line's published result.