{"slug": "rwkv-p-a-neural-network-beat-both-anki-algorithms-for-9999-of-9999-users", "title": "RWKV-P, a neural network, beat both Anki algorithms for 9,999 of 9,999 users", "summary": "A neural network called RWKV-P, built by a programmer known as Alex (1DWalker on GitHub), produced better recall-probability predictions than both of Anki's spaced-repetition algorithms for 9,999 of 9,999 users against SM-2 and FSRS-6 and 9,997 of 9,999 against FSRS-7 in the srs-benchmark test run by the volunteers who created FSRS, scoring 0.28 log loss versus 0.34 for FSRS-7 across roughly 350 million scored reviews. RWKV-P uses 2.8 million parameters and reads each user's full review history, including answer time, day of week, daily review count and deck, while FSRS-7 relies on 34 per-user weights. Official Anki does not include RWKV-P; Anki's creator said in 2025 that adding it did not sound practical \"in the near future,\" so it exists only in unofficial Anki versions.", "body_md": "# RWKV-P, a neural network, beat both Anki algorithms for 9,999 of 9,999 users\n\nAnki is a flashcard app that decides when to show you each card again. By default it uses SM-2, a formula from 1987. You can switch to FSRS, which guesses how likely you are to remember each card before you see it. Cards you're about to forget come up today. Cards you know well wait a few weeks.\n\nA neural network called RWKV-P (RWKV-Instant on the benchmark's chart) makes those guesses better than FSRS does. In a public test on 9,999 Anki users, its guesses were closer for 100.0% of them. On 19 November, an Anki user [posted the chart on X](https://x.com/PantsuThief/status/2123277729235939752), saying it \"destroys any other algorithm\" ever tried. Official Anki doesn't include RWKV-P. In 2025, Anki's creator said adding it didn't sound practical \"in the near future\".\n\n## It beat the newest FSRS for 9,997 of 9,999 users\n\nThe test is called [srs-benchmark](https://github.com/open-spaced-repetition/srs-benchmark), run by the volunteers who made FSRS. They collected the review histories of about 10,000 Anki users. Before each review, every algorithm in the test guessed how likely the person was to get the card right. Then they checked each guess against the button the person pressed. About 350 million reviews were scored this way.\n\n| Algorithm | Users where RWKV-P guessed better | \n|---|---|\n| SM-2, Anki's original | 9,999 of 9,999 | \n| FSRS-6, in Anki today | 9,999 of 9,999 | \n| FSRS-7, the newest FSRS | 9,997 of 9,999 | \n\nAgainst FSRS-7, 9,997 of 9,999 rounds to 100.0%. RWKV-P lost for the other two users. The chart on X comes from a May 2025 run of the same test. It shows the same 100.0% against FSRS-6, the version you can turn on in Anki today.\n\nBetter here means a lower log loss, a standard score for probability guesses. A confident guess that turns out wrong costs the most. Over the whole test, RWKV-P scored 0.28 and FSRS-7 scored 0.34. That gap is about three times what FSRS-7 gains from being tuned to each user. With its default settings, FSRS-7 scores 0.36.\n\n## RWKV is built from 2.8 million numbers, FSRS-7 from 34\n\nFSRS-7 has 34 numbers called weights, tuned to each user. It looks at one card at a time: when you reviewed it and which button you pressed.\n\nRWKV was built by a programmer called Alex (1DWalker on GitHub). It reads your whole history across every card. It also sees things FSRS never gets, like how long you took to answer, the day of the week, how many reviews you've already done today, and which deck a card is in. It was trained on 5,000 users and tested on 5,000 others, then the two groups were swapped. Everyone gets the same model, with no tuning per user.\n\nExpertium, who developed FSRS-7, gave an example on the Anki forums. A user is in a good mood and keeps pressing Good. Then the mood gets worse and they start pressing Again. In Expertium's words, \"FSRS cannot take advantage of this pattern, RWKV-P can.\"\n\n## No due dates\n\nUnlike Anki, with RWKV-P there's no list of cards due today. The model keeps recalculating your chance of remembering every card, and showing you ones that have dropped below the target you set. It makes each guess right before a review, using everything up to that moment, including cards you answered a minute earlier. Expertium put it like this: \"Every card is always due, you choose desired retention which acts as a threshold for which cards are shown.\"\n\nExpertium said this mode \"might as well be called 'anarchy'\". In FSRS, your chance of remembering a card only goes down between reviews. In RWKV-P, it can go up while you're not studying. In Expertium's words, \"you won't know whether next time you see this card will be tomorrow or in 10 years.\"\n\nThe same model has a second mode called RWKV-Curve, which gives due dates. It's less accurate, but it still beat FSRS-7 for 88.7% of users.\n\n## It's only in unofficial versions of Anki\n\nThe forum thread about it started in May 2025. Expertium asked Alex about building RWKV into Anki. As Expertium relayed it, Alex \"doesn't want to spend 6-12 months doing Anki development stuff\". Ten days later, Damien Elmes, who created Anki, replied: \"It does not sound practical to integrate in the near future, and we still have our hands full dealing with FSRS.\" Expertium had written earlier in the thread that \"development efforts are better spent elsewhere for now\". FSRS itself has been in Anki since 2023, still switched off by default.\n\nThis summer, Expertium worked on making RWKV small enough for a normal computer. A version in an unofficial Anki needed about 1.8 GB of extra space for an average user. Expertium got a test version under 1 MB. Speed was the other problem. One tester's first load took about 70 minutes. The largest collection in the dataset has 1.2 million cards. Expertium's comment: \"Don't ask me how someone has 1.2M active cards, I have no freakin' clue.\"\n\nIn September, Expertium was building a separate version of Anki called Clanki, and said RWKV would be ready \"by the end of October optimistically\". On 18 October, Expertium wrote: \"I have decided to stop working on RWKV and Clanki\", adding \"I'll try again in 6-12 months\". A developer called JSchoreels had already added RWKV to their own version of Anki in June and kept working on it. It has both RWKV and FSRS-7, with a switch between them.\n\nIf you want to try RWKV now, two unofficial versions of Anki include it, along with FSRS-7: [JSchoreels' build](https://github.com/JSchoreels/anki/releases) for desktop, and the [kuma3 app](https://github.com/Pedrubik2000/kuma3-anki-app/releases) for Android, which Expertium uses for reviews. Neither comes from the Anki team. Both run JSchoreels' versions of RWKV, which are smaller than the one in the test.\n\n## The Anki forum is arguing about it\n\nOn 15 November, Abdo, a member of the Anki team, wrote that it's \"too early\" for more schedulers, \"especially when FSRS is 'good enough'\", and suggested letting people add their own schedulers as add-ons. Two days later, in a post marked as personal opinion, Abdo added that Anki \"has to be conservative here precisely because the scheduler needs a lot of thought\".\n\nA few users weighed in. One called the unofficial versions \"vibecoded to the extreme\", meaning written mostly by AI. Another wrote \"I am off to JSchoreels' fork, ciao\". A third wanted research backed by peer-reviewed papers before anyone changed Anki's scheduler. Expertium replied that FSRS is \"barely mentioned in academic papers\" either. \"It's not the degree that matters, it's performance on the 10k dataset.\"\n\n## Nobody knows yet if it means fewer reviews\n\nBetter guesses should help a scheduler pick better times to show you a card. The test doesn't measure whether people who use RWKV remember more or study less. Expertium said as much: \"we have no way of converting ML metrics into reviews/day.\"\n\n## Not switching to it yet\n\nI ran the benchmark's RWKV-P on my own reviews in learnfromx, my Japanese flashcard app (2,033 reviews of 884 cards). RWKV-P scored 0.42 and FSRS-6, tuned to my reviews, scored 0.45. That's about a third of the gap in the big test. I hadn't been tracking how long I take to answer, so I gave RWKV-P the same answer time for every review, which might explain the smaller gap. I've started recording answer times, we will see in a few weeks if it's worth switching.\n\n## Sources\n\n- [srs-benchmark](https://github.com/open-spaced-repetition/srs-benchmark) : results, the chart, and what each model sees\n- [Anki forum thread on moving to a neural network](https://forums.ankiweb.net/t/suggestion-to-damien-migrating-to-an-nn-neural-net/60830) : all the quotes above\n- [The post on X](https://x.com/PantsuThief/status/2123277729235939752)", "url": "https://wpnews.pro/news/rwkv-p-a-neural-network-beat-both-anki-algorithms-for-9999-of-9999-users", "canonical_source": "https://learnfromx.com/research/rwkv-anki-fsrs", "published_at": "2026-11-20 14:04:11+00:00", "updated_at": "2026-11-20 14:15:27.989537+00:00", "lang": "en", "topics": ["machine-learning", "neural-networks", "ai-research"], "entities": ["RWKV-P", "Anki", "FSRS", "FSRS-7", "SM-2", "Alex (1DWalker)", "Expertium", "srs-benchmark"], "also_reported_by": [], "takeaway": "RWKV-P's whole-history model with 2.8 million parameters reached 0.28 log loss versus 0.34 for FSRS-7 (0.36 at default settings) on the srs-benchmark, beating FSRS-7 for 9,997 of 9,999 users without any per-user tuning.", "recommendation": {"advice": "Benchmark memory or prediction models on held-out users with log loss rather than per-user tuning, and swap train/test groups to confirm gains; RWKV-P's 0.28 beat FSRS-7's 0.34 this way.", "playbook": "evals-testing", "playbook_url": "https://wpnews.pro/playbooks/evals-testing"}, "alternates": {"html": "https://wpnews.pro/news/rwkv-p-a-neural-network-beat-both-anki-algorithms-for-9999-of-9999-users", "markdown": "https://wpnews.pro/news/rwkv-p-a-neural-network-beat-both-anki-algorithms-for-9999-of-9999-users.md", "text": "https://wpnews.pro/news/rwkv-p-a-neural-network-beat-both-anki-algorithms-for-9999-of-9999-users.txt", "jsonld": "https://wpnews.pro/news/rwkv-p-a-neural-network-beat-both-anki-algorithms-for-9999-of-9999-users.jsonld"}}