{"full_name":"jaredpalmer/kev","owner":"jaredpalmer","name":"kev","url":"https://github.com/jaredpalmer/kev","description":"Jev-like family of decision models built on top of Qwen3.5/3.8 you can train and run on your own","homepage":"","topics":[],"language":"Python","license":"Apache-2.0","avatar":"https://avatars.githubusercontent.com/u/4060187?u=97bb33a7ca5081fcf173c14207e04cbc6c650650&v=4","created_at":"2026-09-17T20:49:39Z","pushed_at":"2026-11-03T00:46:58Z","metrics":{"stars":9928,"forks":677,"watchers":33,"issues":47,"prs":231,"releases":3,"issues_open":null,"issues_closed":null,"prs_open":null,"prs_closed":null,"prs_merged":null,"discussions":0},"growth":{"day":{"stars":22,"forks":1,"base_day":"2026-11-09"},"week":{"stars":161,"forks":9,"base_day":"2026-11-03"},"month":{"stars":1429,"forks":117,"base_day":"2026-10-11"}},"history":{"dates":["2026-09-30","2026-10-01","2026-10-02","2026-10-03","2026-10-04","2026-10-05","2026-10-06","2026-10-07","2026-10-08","2026-10-09","2026-10-10","2026-10-11","2026-10-12","2026-10-13","2026-10-14","2026-10-15","2026-10-16","2026-10-17","2026-10-18","2026-10-19","2026-10-20","2026-10-21","2026-10-22","2026-10-23","2026-10-24","2026-10-25","2026-10-26","2026-10-27","2026-10-28","2026-10-29","2026-10-30","2026-10-31","2026-11-01","2026-11-02","2026-11-03","2026-11-04","2026-11-09","2026-11-10"],"stars":[2376,3898,5798,6697,6968,7181,7619,7861,8174,8303,8364,8499,8582,8644,8734,8824,8895,9001,9065,9121,9184,9237,9262,9319,9371,9418,9453,9478,9508,9566,9612,9656,9690,9717,9767,9796,9906,9928],"forks":[122,193,314,383,412,428,473,495,520,534,546,560,562,568,571,578,585,599,608,612,619,623,624,627,634,639,640,644,647,653,662,667,666,666,668,669,676,677],"watchers":[8,15,21,26,25,25,25,26,28,29,29,28,27,27,29,30,30,33,32,32,32,32,32,32,32,32,32,32,32,32,32,32,33,33,33,33,33,33],"issues_open":[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null],"prs_open":[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null]},"boards":{"monthly_rate":109,"new":51},"same_owner":[],"issues":[{"number":71,"title":"Embedded Device Support (Jetson) and Profiling","comments":6,"created_at":"2026-09-23T01:22:33Z"},{"number":107,"title":"Serving: remaining work after batching and fused kernels (#93, #97)","comments":4,"created_at":"2026-09-24T13:18:02Z"},{"number":28,"title":"Getting kev to work with images ?","comments":3,"created_at":"2026-09-22T04:13:23Z"},{"number":67,"title":"LiteRT / Android export for Kev-0.8B","comments":2,"created_at":"2026-09-22T23:31:50Z"},{"number":48,"title":"Long-context recipe (v10): Kev loses accuracy on documents longer than its 384-token training context","comments":2,"created_at":"2026-09-22T17:24:58Z"},{"number":42,"title":"Have you considered using OpenBMB's MiniCPM5-2B model to train kev?","comments":2,"created_at":"2026-09-22T16:22:09Z"},{"number":255,"title":"`--device auto` on Apple Silicon runs the same MPS probe twice before falling back to CPU","comments":1,"created_at":"2026-10-22T23:55:18Z"},{"number":155,"title":"Community Kev-2B on Qwen3.5-2B-Base: Kev-0.8B release recipe, test results and two small findings","comments":1,"created_at":"2026-09-26T12:03:29Z"}],"discussions":[],"latest_release":{"name":"Kev 1.0","tag":"kev-1.0","published_at":"2026-10-01T17:25:27Z","body":"Kev 1.0 is the first versioned release of the whole Kev family: four decision models that read a document and a set of typed questions and return calibrated probabilities over the options, in one forward pass, behind TypeSafe's System One API. Nothing in it is newly trained. It pins the checkpoints, model cards, evaluation suites and serving code that the next generation of Kev is measured against, with the same tag (`v1.0`) on every Hub repo.\n\n## What's in 1.0\n\n| Model | Hub repo | Weights revision | Form | Base | Temperature | Validated context |\n|---|---|---|---|---|---|---|\n| Kev-0.8B | [`jaredpalmer/kev-0.8b`](https://huggingface.co/jaredpalmer/kev-0.8b) | `9a45d25e` | LoRA adapter + head | Qwen3.5-0.8B-Base (Apache-2.0) | 2.35 | 8,192 tokens |\n| Kev-4B | [`jaredpalmer/kev-4b`](https://huggingface.co/jaredpalmer/kev-4b) | `139fdd94` | LoRA adapter + head | Qwen3.5-4B-Base (Apache-2.0) | 2.41 | 8,192 tokens |\n| Kev-9B (v2) | [`jaredpalmer/kev-9b`](https://huggingface.co/jaredpalmer/kev-9b) | `b5d8c18e` | LoRA adapter + head | Qwen3.5-9B-Base (Apache-2.0) | 2.19 | 8,192 tokens |\n| Kev-27B (v2) | [`jaredpalmer/kev-27b`](https://huggingface.co/jaredpalmer/kev-27b) | `28be62e9` | full bf16 weights (51 GB) + head | Qwen3.8-27B, post-trained (Apache-2.0) | 1.32 | 65,536 tokens |\n\nHeadline numbers (fp32 evaluation path, each model at its shipped temperature; the transfer-v4 test is locked and was read once per model):\n\n| | Kev-0.8B | Kev-4B | Kev-9B | Kev-27B | Jev |\n|---|---|---|---|---|---|\n| Held-out datasets: breadth-v1 test, chance-corrected index | 23.3 | 38.0 | 41.0 | 52.3 | 54.0 |\n| Out-of-domain: transfer-v4 development accuracy | 0.648 | 0.817 | 0.820 | 0.851 | 0.857 |\n| Out-of-domain: transfer-v4 locked test accuracy / Brier | 0.697 / 0.397 | 0.838 / 0.224 | 0.852 / 0.199 | 0.889 / 0.154 | – |\n| Skills: hard-v1 test | 0.665 | 0.803 | 0.834 | 0.918 | – |\n| Developer tooling: devtools-v1 test, all sources | 0.637 | 0.756 | 0.791 | 0.790 | – |\n| Real documents: documents-v1 test | 0.851 | 0.903 | 0.900 | 0.908 | – |\n| MMLU-Pro (transfer-v9 development) | 0.230 | 0.565 | 0.590 | 0.675 | 0.840 |\n\nhard-v1, devtools-v1 and documents-v1 have training splits that every Kev trained on: those rows measure held-out items of trained families, not transfer. The breadth-v1 and transfer-v4 rows are datasets no Kev trained on. Jev was read on the development partitions and on breadth-v1 test only. Every number traces to a committed report through `docs/claims.json`; the model cards (`docs/model-cards/`) have the rest, with intervals.\n\n## What changed since the last family release\n\nMeasured from the GitHub release `kev-family` as first assembled for the current family on 2026-09-24 (Kev-27B v1, Kev-9B v1, and the same Kev-4B and Kev-0.8B as here). Its 2026-09-30 updates (Kev-27B v2, Kev-9B v2) are listed here too, since 1.0 is where they become part of a versioned release.\n\n- **Kev-27B v2: full weights.** Every weight of Qwen3.8-27B fine-tuned for one epoch on a 145,840-record corpus, then averaged 0.85 / 0.15 with v1. Against v1 on test: held-out datasets +1.2 pp [+0.3, +2.2], held-out task families +5.3 [+3.7, +6.8], skills, tooling and documents +8.9 [+7.5, +10.3]; locked out-of-domain test 0.889 against 0.896, Brier 0.154 against 0.160. Worse and overconfident on long contracts (CUAD ECE 0.053 against 0.007). v1 is at `jaredpalmer/kev-27b@v1-lora`.\n- **Kev-9B v2.** v1 plus one epoch on the documents and skills data Kev-4B and Kev-0.8B already had. Against v1 on test: hard-v1 + devtools-v1 +18.7 pp [+16.7, +20.8], documents-v1 +7.1 [+4.7, +9.2]; level on the locked out-of-domain test (0.852 both) with Brier 0.199 against 0.224. v1 is at `jaredpalmer/kev-9b@v1`.\n- **No silent truncation.** The server used to cut a state longer than its limit without saying so. It now refuses a state over 65,536 tokens with a 422 that names the token count and the limit; `KEV_TRUNCATE_STATES=1` opts back into truncation, and every response from such a "},"readme":true,"readme_zh":false,"news":[{"title":"Kev: Tiny Jev-like family of decision models built on top of Qwen3.5","url":"https://github.com/jaredpalmer/kev/tree/main","hn":"https://news.ycombinator.com/item?id=49783999","source":"Hacker News","lang":"en","date":"2026-09-21","points":463,"comments":210},{"title":"kev: Jev-like model built on Qwen2.5-0.5B","url":"https://github.com/jaredpalmer/kev","hn":"https://news.ycombinator.com/item?id=49766295","source":"Hacker News","lang":"en","date":"2026-09-19","points":3,"comments":0},{"title":"Jared Palmer Ports Kev to Qwen3.5 for Roughly $95 in H100","url":"https://aiweekly.co/alerts/jared-palmer-ports-kev-to-qwen35-for-roughly-95-in-h100","source":"aiweekly","lang":"en","date":"2026-09-21"},{"title":"Exploited Vulnerabilities Can Take Months to Make KEV List","url":"https://www.darkreading.com/vulnerabilities-threats/exploited-vulnerabilities-can-take-months-to-make-kev-list","source":"Dark Reading","lang":"en","date":"2023-11-19"},{"title":"Node2Shell: Node.js Brotli Flaw CVE-2026-111497 Added to CISA KEV, Upgrade to 22.23.4, 24.22.1 or 26.12.1","url":"https://news.google.com/rss/articles/CBMijwFBVV95cUxOOGFRRjNjbEFxMWRhRmlhX25wSS02Y1dISVRaRXJaTGJKbDdMZW5yRTF3bE9RRGhrVVNqNE9kZjloeGhaV3Q2ZXVRS1F3bFNqQ1hnM1l0TDc2SFE0Y0FxVW5sc3VjajYzem9JU0d3LXJNSWlUS1lNc0F3Z1dqRjNFWDZjRHdBMkw0bkFUeEE5Yw?oc=5","source":"Rescana","lang":"en","date":"2026-10-23"},{"title":"Nimble v3 与 Kev：一个自带配方，一个则没有","url":"https://news.google.com/rss/articles/CBMiY0FVX3lxTFBEQ0NKMFJuemFQY3lKWDB5SGQwT2JyN25lRGp4c3JWeGJrMHUwQTRUVlFHTEc4QXUzTjBybDNweldXUmNDcHRQQVBaUEdROWN6S0VONVhnZUdDbFVWYkc5UUppNA?oc=5","source":"OrcaRouter","lang":"zh","date":"2026-10-13"},{"title":"Laya vs Kev：以检查点对比训练配方","url":"https://news.google.com/rss/articles/CBMiXEFVX3lxTFBBZUhUSzZWSUF3VFJjbW9peGRZOXVSVU5naEE5WTNzR2JjOTBqWlBGaEJTbWhVRkJ2NlVxVnFFRlp3Z3FRWHdJWm5nQ05BYXhReVQwN0RPMGRjV0JV?oc=5","source":"OrcaRouter","lang":"zh","date":"2026-09-23"},{"title":"Jev模型：又快又便宜，消灭幻觉，轻松打败顶流LLM","url":"https://news.google.com/rss/articles/CBMiVkFVX3lxTFBRa0NFVUJuSTAydEdYOWNMMy1xaUg5RnZtdDlTcnFOMGEtVGk0dUVmWVcxb08xY29IYjF4Q09VSjc2U2dNRmE4dTk1VXJrWnZBZHI5eGJR?oc=5","source":"腾讯新闻","lang":"zh","date":"2026-09-19"},{"title":"ownCloud Flaw Exploited to Steal Nuclear Records From Philippine Research Body","url":"https://news.google.com/rss/articles/CBMifkFVX3lxTE9GVllVUG5uR3RfUDNiTHJQUU03TUl0OU0yeHNUcEFzU2E3SnUyR3pwQTU3ZEhPdlBDX1VVWGtIQUk5OXVKVmJfN2FtLXVaZUlRdnF2ckdiRnB5TmVrZHNQOWRxa2lGb0pReXd6Yl9tOUJsMTdMd0xJNWtzSklyZw?oc=5","source":"The Hacker News","lang":"en","date":"2026-08-28"},{"title":"Critical Command Injection Vulnerability in Snowflake snowflake-connector-net GitHub Actions Exposes Jira Credentials","url":"https://news.google.com/rss/articles/CBMi3wFBVV95cUxOcnItTGtOYjRoMzczRDJPYldiNWxGelowVXVRenYzdERDQTRuNFh0OFVjMlJfNHM5VHFHUk1xM1d5MWNtSnl6bjR4bGZPeVZfWklaaFVzbkh3V0l5anhyVHY4NUVSZGhfRDdpRXJWX3p3dWpUSV9MMHlOWGRENUhIX2dEbzdVN2hxOFNlRnZEeDZIUjAwdVl4eWtTVU1JVkFTeEpheFE4TjdWaTVDZVg5YmJCN0MtdFFPMlRST2pIYmJYZGJOenBmYnlVUXZ1WFNNQmpkdWZBSlg3TGsxS1Jv?oc=5","source":"Rescana","lang":"en","date":"2026-08-18"},{"title":"CISA warns of actively exploited Git code execution flaw","url":"https://news.google.com/rss/articles/CBMipwFBVV95cUxQdUh0Yjc1bTBYbVExRzhWbHFRZmw0WmpodW42STFOYWdXcVZMMWgzcXhuaTR1T3FIbnNZeXZlOGlteUk5VktlX2RiaENuVDAtZWxGNU1aYTdZT0VVMHV3b0tHZ2dYLTUxUWFITE0wRWtPTHg5Rjk4dklNWUJaRXhxazJmclBNa0pvYy1haGFLZmliRlQ4LUNIaFVQY2x6UmN4NFdpV1VTb9IBrAFBVV95cUxNQTQxSWtENk9BdnMtQ19uYjJjZzAyVU9DQlc5Yll1aDgwM1ZCbENGQ2pGSUpNa1E4Sm5qNEhLOHZJYUlHZWliM0RrNTJVV2lCTkhjMnZiT25nanBDYVZsRDE2Z2pPMTV1eEpuME5XQ3NySTZYY3JrZFRIb0g4aExCdEpnbE5KUk5NYTJwRE9iejlmYjViMlZXRVJhb0d2Wl9QM0RBSjljenZQSHRX?oc=5","source":"BleepingComputer","lang":"en","date":"2025-08-26"},{"title":"CISA Shifts Alert Distribution Strategy to Email, Social Media","url":"https://news.google.com/rss/articles/CBMiggFBVV95cUxNYU5mTDk0ZDNINTZMcGFlV0FHU0RlcDlBRC1NSC1ObmxUZmF6aEdEeWVGTjZ3SWVsSzloVFVyRDMtQ1UtR2lTYUpFMk5YME41a2pJR01vRWpNZFN5M2V1Q09sM0lWLS1DR0N0Y1o4a1Y3bjkyNUNMY1Jkd3kxSFZ4VTZ3?oc=5","source":"Infosecurity Magazine","lang":"en","date":"2025-05-13"},{"title":"盘点：20个大幅提高效率的开源网络安全工具","url":"https://news.google.com/rss/articles/CBMiTkFVX3lxTE9sU0dPWmdrbFNTb1FLOFF5cHBOLXc4UkhTZUJNTTVRdkV5S01IWmRnUnl1QUs3LXJacWE5UWZoY3JyYkk0QUJlU2VrbVNIQQ?oc=5","source":"安全内参","lang":"zh","date":"2024-03-26"}],"stale":false,"snapshot_at":"2026-11-10T09:35:25Z","history_start":"2026-09-06"}