Post
CommunityJev Media & Talks
Jev plays chess
Maxim Saplin ran TypeSafe Jev on LLM Chess by turning each position into one Choice question over the legal moves. jev-latest landed around rank 59 at Elo 242.9, next to qwen3.6-27b and o4-mini-medium, at roughly $0.0015 and 36 seconds per game.
- Format
- Post
- Reading time
- 3 min
- Published by
- Community
- Author
- Maxim Saplin
- Last verified
- Oct 2026
Used inJev for Gaming
Highlights
- Each ply is one Jev Choice: state is the FEN plus side to move, and the options are all legal UCI moves with SAN text in the criteria, max 255.
- jev-latest sits around #59 at Elo 242.9 +/- 117.5, next to qwen3.6-27b (270.0) and o4-mini-medium (240.3) in the post's table.
- 119 tokens, $0.0015 and 35.6s per game over 80 games, versus $1.11 to $1.46 and minutes to hours for its Elo neighbours.
- Game duration 100% with full games and no illegal moves; draws versus Dragon held at 5 of 10 games at each of levels L1, L2 and L3.
- Published pricing in the post: $0.042 per 1M input tokens and free output; harness is maxim-saplin/llm_chess on the feat/typesafe-jev-player branch.
Watch out
Copy reviewed against the full dev.to post on 2026-10-15. The rank, Elo and cost numbers come from the author's LLM Chess leaderboard at the time of writing rather than an independent rerun, and the post notes TypeSafe is publicly against benchmark scoreboards, so treat this as one outside eval of picking a move from a fixed list.
Questions about this entry
Short answers drawn from the linked project or page.
Maxim Saplin ran TypeSafe Jev, which is not a chat model, through LLM Chess: for each position he gave Jev the legal moves as the options of one Choice question and asked it to pick one. The post compares the result against reasoning models on the same leaderboard.
The harness hands chat models an agentic protocol (get_current_board, get_legal_moves, make_move), but Jev only answers typed questions. So each ply becomes one Choice: state is the FEN plus side to move, and the options are all legal UCI moves, with SAN text in the criteria, capped at 255 options.
Around rank 59 at Elo 242.9, between qwen3.6-27b at 270.0 and o4-mini-medium at 240.3 in the table shown. Game duration was 100% across 80 games, which the author reads as perfect protocol following rather than early deaths on illegal moves.
119 tokens, $0.0015 and 35.6 seconds per game on average; pricing in the post is $0.042 per 1M input tokens with output free, which works out to about $0.00002 per move. The 80 games cost roughly $0.12 in total.
A next-character loop over a-z plus <eos>, one API call per letter, produced almost-words: it answered bue for blue and helhoh for hel..., though starting from pari... it finished paris. The author's takeaway is that Jev fits closed option lists far better than spelling English character by character.
Against the Dragon engine it drew 5 of 10 games at each of levels L1, L2 and L3, a flat 50% draw rate across three levels. Chat models usually draw less as the engine strengthens, so the flat rate helped the Elo rather than hurting it.
The harness is maxim-saplin/llm_chess on the feat/typesafe-jev-player branch, with results on the LLM Chess leaderboard. The post notes TypeSafe is openly anti-benchmaxxing and skips a public scoreboard, so this is one outside eval, not a vendor result, and not a statement about Jev on any other task.
More Jev resources like this
A Python benchmark that evaluates LLMs' reasoning and instruction-following by making them play chess against a random player or the Komodo Dragon engine in multi-turn dialogs, separating chess skill from protocol durability.
Jev Benchmarks & Evaluations#community#catalog-import#madewithjev
Community
maxim-saplin/llm_chessPython
maxim-saplin/llm_chessPythonA browser chess coach where Stockfish (WASM) owns the truth about the position and Jev makes the coaching judgment after every move, with templates and a small writing model supplying the words. Status: Wave 0 scaffold and Jev feasibility spike.
Jev Demos & Experiments#community#catalog-import#madewithjev
Community
joellewis/game-coachTypeScript
joellewis/game-coachTypeScriptA playable Three.js driving simulator where Jev-powered autopilot picks steering and speed up to four times per second: compact tables of eligible paths, traffic and road boundaries go to Jev while geometry and control math stay local.
Jev Demos & Experiments#community#catalog-import#cobanov
Community
standardagents/jevpilotJavaScript
standardagents/jevpilotJavaScriptA worked example of using Jev as an LLM judge: ten synthetic frozen refund-support traces are scored by three yes/no criteria and one ordered helpfulness rubric in a single Jev request and recorded as a Comet Opik experiment.
Jev Benchmarks & Evaluations#community#catalog-import#madewithjev
Community
patchy631/jev-as-judgePython
patchy631/jev-as-judgePythonA Three.js first-person garden game where a glowing fly hunts goblins and TypeSafe's Jev steers: about five times per second it answers four typed questions on turning, sidestepping trunks, climbing and diving, and every flight is logged.
Jev Demos & Experiments#community#catalog-import#madewithjev
Community
webdevcody/jevs-flyJavaScript
webdevcody/jevs-flyJavaScriptProphetLab's Texas Hold'em benchmark and playground for decision models: separate cash and sit-and-go leaderboards, hand replays, a Jev-backed advisor, and bring-your-own-agent tables, all with virtual chips.
Jev Benchmarks & Evaluations#community#catalog-import#madewithjev
Community
prophetlab/jevpokerbenchPython
prophetlab/jevpokerbenchPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.