Released January 31, 2026 — 46 ELO stronger than its predecessor. Free, open source, and available right now.
Stockfish 18 launched on January 31, 2026 with a gain of +46 ELO over version 17. The improvement comes from a redesigned neural network called SFNNv10 — not faster calculation, but significantly deeper positional understanding. It remains free and open source under the GPLv3 license.
Our platform already runs the latest engine across all tools: the chess calculator, game analyzer, and AI bots. You can use it right now, directly in your browser, with no download required.
Our Chess Tools
Here's a detailed look at what changed between the two versions:
| Feature | Stockfish 17 | Stockfish 18 |
|---|---|---|
| ELO Rating (CCRL) | ~3600 | ~3650 (+46) |
| Neural Network | SFNNv8 | SFNNv10 |
| Threat Inputs | No | Yes (new) |
| Shared Memory | No | Yes (new) |
| Correction History | Basic | Enhanced |
| Chess960 Support | Yes | Yes (improved) |
| Head-to-Head Win Ratio | — | 4:1 vs SF17 |
| UCI_Elo Range | 1320–3190 | 1320–3190 |
The +46 ELO gain doesn't mean the engine searches faster — it means the evaluation is smarter. Positions that version 17 assessed as equal, the new engine correctly identifies as having subtle advantages. In direct matches, it wins four games for every one it loses to its predecessor.
The engine's strength varies depending on the time control. Here are the CCRL-published ratings:
| Time Control | ELO Rating | Source |
|---|---|---|
| Standard (40/15) | ~3650 | CCRL |
| Bullet | ~4060 | CCRL Blitz |
| Blitz | ~3999 | CCRL Blitz |
| Rapid | ~4050 | CCRL 40/2 |
To put these numbers in perspective:
The strongest chess engine in history. Breaks 4000 in faster time controls.
The highest FIDE rating ever achieved by a human — nearly 800 points below the engine.
Even elite human players would lose virtually 100% of games against the engine at full strength.
Yes — and it's not close. An 800-point ELO gap translates to a win probability above 99% for the higher-rated side. Carlsen winning a game would be roughly equivalent to a 1400-rated club player beating Carlsen himself.
Chess engines surpassed the best humans around 2006 when Deep Fritz beat Kramnik, and the gap has only widened since. The version 18 release represents the largest lead an engine has ever held over human players.
Four changes in this release directly translate to stronger play. Here's what each one does and why it matters:
The biggest single change. SFNNv10 is the 10th generation of the NNUE (Efficiently Updatable Neural Network) architecture, trained on over 100 billion positions using Leela Chess Zero evaluation data. NNUE lets the engine evaluate positions with human-like intuition while maintaining lightning-fast alpha-beta search speed.
The key innovation in SFNNv10. The neural network now directly "sees" which pieces attack and defend each square. Previously, the engine discovered threats only through deeper search. Now threats are part of the evaluation input — better tactical "intuition" without extra computation time.
The engine now dynamically adjusts evaluations during search based on what it learns in real time. If it discovers that a position type was consistently misevaluated, Correction History applies adjustments on the fly. This particularly improves fortress detection and drawn endgame evaluation.
Multiple engine processes can now share neural network weights in memory. This matters for cloud analysis platforms and servers running multiple instances — it cuts memory usage dramatically without affecting analysis quality.
ELO gains at the engine level work differently than for human players. Here's the practical impact:
The engine combines two techniques to find the best chess moves:
Alpha-Beta Search: It looks ahead many moves by building a tree of possible positions. Alpha-beta pruning cuts branches that can't improve on what's already found. At depth 20, the engine effectively examines 20+ moves ahead while evaluating millions of positions per second.
NNUE Neural Network: At each position in the search tree, the engine needs to evaluate "how good is this?" That's where SFNNv10 comes in. Inspired by AlphaZero's approach, it learns patterns from billions of games. The key advantage over traditional neural networks is speed: NNUE updates incrementally as pieces move, making it fast enough for deep search.
The combination is what dominates: Deep tactical calculation from alpha-beta search, plus positional understanding from the neural network. Neither approach alone is as strong as both together — which is why this hybrid architecture beats pure neural network engines like Leela Chess Zero.
Leela Chess Zero (Lc0) is the main rival — a pure neural network engine inspired by DeepMind's AlphaZero. The two use fundamentally different approaches:
| Aspect | Stockfish 18 | Leela Chess Zero |
|---|---|---|
| Search Method | Alpha-beta + NNUE | Monte Carlo Tree Search (MCTS) |
| Best Hardware | CPU (any modern processor) | GPU (expensive, high-end) |
| CCRL Rating | ~3650 | ~3580 |
| Runs in Browser | Yes (WebAssembly) | No (needs GPU) |
| Price | Free | Free (but needs GPU) |
For most chess players, the Stockfish engine is the better practical choice. It runs on any computer or phone, works in the browser via WebAssembly, and is consistently rated higher. Lc0 produces creative, human-like play but requires expensive GPU hardware.
AlphaZero's famous 2017 match against Stockfish 8 happened nearly a decade ago — against a version without any neural network evaluation. Since then, Stockfish adopted the same neural network approach that made AlphaZero revolutionary and combined it with superior alpha-beta search.
AlphaZero was never updated after 2017, while the Stockfish project has gained over 500+ ELO since that match. The version that AlphaZero defeated would lose to today's engine by an even larger margin than it lost to AlphaZero. The chess engine landscape has changed fundamentally.
Beyond Lc0, several strong engines compete on the CCRL rating list. Here's the current top tier:
| Engine | CCRL ELO | Type |
|---|---|---|
| Stockfish 18 | ~3650 | Alpha-beta + NNUE |
| Torch | ~3635 | Alpha-beta + NNUE |
| Obsidian | ~3631 | Alpha-beta + NNUE |
| Berserk | ~3610 | Alpha-beta + NNUE |
| Leela Chess Zero | ~3580 | MCTS + Neural Network |
| Komodo Dragon | ~3560 | Alpha-beta + NNUE |
The Stockfish project has improved consistently with each major release. Here are the recent milestones from the official blog:
SFNNv10, Threat Inputs, Shared Memory. +46 ELO over version 17.
SFNNv8, improved search, UCI_Elo calibration. +44 ELO over version 16.
Major NNUE improvements, dual-net evaluation. +50 ELO over version 15.
First NNUE neural network integration — the biggest single leap in the project's history.
You don't need to install anything. ChessNextMove runs the engine directly in your browser using WebAssembly:
Set up any position, paste a FEN, or upload a screenshot. Get depth 18+ analysis in seconds. Try it free →
Paste your PGN and get every move classified as brilliant, good, inaccuracy, mistake, or blunder. Analyze a game →
Practice against 11 engine-powered bots calibrated from complete beginner to full strength. Choose a bot →
Running the engine locally in a chess GUI (Arena, SCID, ChessBase)? Here are optimal settings:
Set to roughly half your available RAM. For 8GB RAM, use 4096 MB. For 16GB, use 8192 MB. More hash means better analysis in longer time controls.
Set to your CPU core count. A 6-core processor should use Threads=6. More threads means faster and deeper search.
Depth 20-25 is sufficient for most positions. Depth 30+ helps in complex tactical positions but takes significantly longer. For quick checks, depth 18 already provides strong results.
Range: 1320-3190. Enable UCI_LimitStrength, then set your desired opponent rating. Below 1320, combine Skill Level (0-20) with shallower depth for weaker play.
The engine fully supports Chess960, where the back rank pieces are randomized. The SFNNv10 network handles non-standard starting positions more accurately than previous versions — Threat Inputs evaluate piece relationships regardless of where pieces started.
To enable it, set UCI_Chess960 to true. All UCI-compatible GUIs that support Chess960 will work without additional configuration.
Having access to the strongest engine means nothing if you don't use it effectively. Here's a level-by-level approach:
Play against our low-rated bots (Wobble at 400, Blinky at 800). After each game, use the game analyzer to find your blunders. Focus on eliminating one-move mistakes before worrying about strategy.
Use the calculator to analyze critical positions from your games. Compare your moves to the engine's suggestions and understand why it prefers different lines. Focus on inaccuracies, not just blunders. Try our ELO test to benchmark your current level.
Study the top 3 candidate moves in complex positions. At this level, the gap between the best and second-best move often reveals subtle positional concepts. Use deep analysis (depth 25+) for opening preparation and endgame study.
Many online chess calculators still run Stockfish 11 or even older versions. Upgrading requires building new WebAssembly binaries, compiling server binaries, testing, and redeploying — significant engineering work that most free tools never do.
The gap matters. Stockfish 11 (released 2019) is roughly 300+ ELO weaker than the current version. That's the difference between a strong grandmaster and a world champion. Outdated engines give outdated analysis, especially in complex middlegame positions.
We upgraded within weeks of the January 2026 release. When you use our calculator or analyzer, you're getting the most accurate analysis available anywhere online — for free.
The upgrade brings measurable gains across all time controls — same UCI protocol, same commands, same settings. Everything is backward compatible.
If you use ChessNextMove, there's nothing to do — we already upgraded. Every analysis, bot game, and calculation on our platform uses the latest version.
It was released on January 31, 2026 — the latest version of the world's strongest open-source chess engine.
Approximately 46 ELO points stronger in internal testing, achieving a 4:1 win ratio in head-to-head matches. The improvement comes from smarter positional understanding, not just faster calculation.
Around 3650 on the CCRL standard rating list. In faster time controls it goes even higher: ~4060 in bullet, ~3999 in blitz, and ~4050 in rapid.
The SFNNv10 neural network with Threat Inputs for better tactical awareness, Shared Memory for multi-process deployments, improved Correction History for more accurate evaluations, Chess960 improvements, and automated training on 100 billion+ positions.
Yes — ChessNextMove runs the engine in your browser via WebAssembly. Use the chess calculator for instant analysis, game analyzer for full reviews, and AI bots for practice. No download or registration needed.
Yes. It tops every major rating list — CCRL, CEGT, and TCEC — ahead of Leela Chess Zero, Komodo Dragon, Torch, Obsidian, and every other engine.
The 10th generation of the NNUE neural network architecture. The key innovation is 'Threat Inputs' — the network directly sees which pieces attack and defend each square, improving tactical awareness without additional computation.
Easily. At ~3650 ELO, the engine is nearly 800 points above Carlsen's peak of 2882. In practical terms, it would win virtually every game against any human player.
Yes, with improved evaluation for non-standard starting positions. The SFNNv10 network handles unusual piece placements more accurately than previous versions.
In head-to-head matches, yes — it consistently outperforms Lc0. It dominates on CPU hardware, while Lc0 is competitive on high-end GPUs. For most players, it is the stronger and more practical choice.
Set Hash to half your RAM (e.g., 4096 MB for 8GB), Threads to your CPU core count, and depth to 20-30. For playing, UCI_Elo (1320-3190) controls strength. Skill Level (0-20) provides finer tuning at lower ratings.
No. You can use it directly in your browser on ChessNextMove via WebAssembly. Just open the chess calculator and start analyzing — no installation or registration required.
The strongest chess engine ever built, running in your browser. Analyze positions, review games, or challenge our AI bots.
Put What You Learned Into Practice
Want more? Try Pro — Try Pro for unlimited AI analysis, 3650 ELO depth & Chrome Extension
