{ "sha": "6a261db7d87b13d76c5197cec636a0a3951afb36", "node_id": "C_kwDOJ0Z1PtoAKDZhMjYxZGI3ZDg3YjEzZDc2YzUxOTdjZWM2MzZhMGEzOTUxYWZiMzY", "commit": { "author": { "name": "Jesse Gross", "email": "jesse@kernel.org", "date": "2026-08-11T18:49:22Z" }, "committer": { "name": "Jesse Gross", "email": "jesse@kernel.org", "date": "2026-08-12T04:47:51Z" }, "message": "api: stop applying repeat_penalty 1.1 to models that don't set one\n\nRequest options are the model's published parameters and the request's\nown options layered over the server defaults, so the default\nrepeat_penalty of 1.1 reaches every model whose parameters leave it\nunset. No maker of the library's current models recommends 1.1: their\ngeneration configs either omit the penalty, meaning 1.0, or pin 1.05.\nllama.cpp dropped the same 1.1 default in 2024; vLLM, SGLang, and\ntransformers apply no penalty. An always-on penalty also distorts\noutput that legitimately repeats tokens, such as code, JSON, and long\nreasoning traces.\n\nThe penalty is especially costly for speculative decoding, where\ndrafts are proposed without it: the penalized target rejects drafted\ntokens and the depth controller backs off. On muse-glimmer 30B (DFlash\non M5 Max, HumanEval) the 1.1 default costs 13-16% of end-to-end\nthroughput at greedy and temperature 1 alike, and drops prose\nacceptance at temperature 0.8 from 0.44 to 0.30. On qwen3.6-35B it\ncuts the mean accepted draft length from 4.3 to 3.5 tokens and makes\nthe controller stop speculating on prose.\n\nDefaulting to 1.0 disables the penalty unless a model's parameters or\nthe request set one. Across the library:\n\n- qwen3, qwen3.6, and qwen3-coder pin their own values (1.0, 1.0, and\n Qwen's recommended 1.05) and are unchanged.\n- Everything else local now matches its maker's no-penalty\n recommendation, including gemma2 through gemma4, muse-glimmer, both\n laguna 2.1 models, qwen3.5 (previously 1.1 stacked on its\n presence_penalty of 1.5), gpt-oss, deepseek-r1 and v3.1, the\n nemotron family, granite4, the mistral and llama3/llama4 families,\n phi4, glm4, llava, and devstral.\n- qwen2.5 recommends 1.05 but ships no parameters, so it moves from\n 1.1 to 1.0 and still needs a parameters layer to conform.\n- Cloud models (kimi-k3, deepseek-v4-flash) never receive these\n defaults.\n\nSmall older models may repeat themselves more without the penalty\nmasking it; the remedy is a per-model parameter, not a penalty applied\nto every model.", "tree": { "sha": "b086e3b0120411d11d1cb8a5a5957d89260efe38", "url": "https://api.github.com/repos/ollama/ollama/git/trees/b086e3b0120411d11d1cb8a5a5957d89260efe38" }, "url": "https://api.github.com/repos/ollama/ollama/git/commits/6a261db7d87b13d76c5197cec636a0a3951afb36", "comment_count": 0, "verification": { "verified": false, "reason": "unsigned", "signature": null, "payload": null, "verified_at": null } }, "url": "https://api.github.com/repos/ollama/ollama/commits/6a261db7d87b13d76c5197cec636a0a3951afb36", "html_url": "https://github.com/ollama/ollama/commit/6a261db7d87b13d76c5197cec636a0a3951afb36", "comments_url": "https://api.github.com/repos/ollama/ollama/commits/6a261db7d87b13d76c5197cec636a0a3951afb36/comments", "author": { "login": "jessegross", "id": 6468499, "node_id": "MDQ6VXNlcjY0Njg0OTk=", "avatar_url": "https://avatars.githubusercontent.com/u/6468499?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jessegross", "html_url": "https://github.com/jessegross", "followers_url": "https://api.github.com/users/jessegross/followers", "following_url": "https://api.github.com/users/jessegross/following{/other_user}", "gists_url": "https://api.github.com/users/jessegross/gists{/gist_id}", "starred_url": "https://api.github.com/users/jessegross/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jessegross/subscriptions", "organizations_url": "https://api.github.com/users/jessegross/orgs", "repos_url": "https://api.github.com/users/jessegross/repos", "events_url": "https://api.github.com/users/jessegross/events{/privacy}", "received_events_url": "https://api.github.com/users/jessegross/received_events", "type": "User", "user_view_type": "public", "site_admin": false }, "committer": { "login": "jessegross", "id": 6468499, "node_id": "MDQ6VXNlcjY0Njg0OTk=", "avatar_url": "https://avatars.githubusercontent.com/u/6468499?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jessegross", "html_url": "https://github.com/jessegross", "followers_url": "https://api.github.com/users/jessegross/followers", "following_url": "https://api.github.com/users/jessegross/following{/other_user}", "gists_url": "https://api.github.com/users/jessegross/gists{/gist_id}", "starred_url": "https://api.github.com/users/jessegross/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jessegross/subscriptions", "organizations_url": "https://api.github.com/users/jessegross/orgs", "repos_url": "https://api.github.com/users/jessegross/repos", "events_url": "https://api.github.com/users/jessegross/events{/privacy}", "received_events_url": "https://api.github.com/users/jessegross/received_events", "type": "User", "user_view_type": "public", "site_admin": false }, "parents": [ { "sha": "948f69330acf96a2310f1b53fdfc211731a386d8", "url": "https://api.github.com/repos/ollama/ollama/commits/948f69330acf96a2310f1b53fdfc211731a386d8", "html_url": "https://github.com/ollama/ollama/commit/948f69330acf96a2310f1b53fdfc211731a386d8" } ], "stats": { "total": 4, "additions": 2, "deletions": 2 }, "files": [ { "sha": "c0e771d4382988a115efa83a63944f3a037eb433", "filename": "api/types.go", "status": "modified", "additions": 1, "deletions": 1, "changes": 2, "blob_url": "https://github.com/ollama/ollama/blob/6a261db7d87b13d76c5197cec636a0a3951afb36/api%2Ftypes.go", "raw_url": "https://github.com/ollama/ollama/raw/6a261db7d87b13d76c5197cec636a0a3951afb36/api%2Ftypes.go", "contents_url": "https://api.github.com/repos/ollama/ollama/contents/api%2Ftypes.go?ref=6a261db7d87b13d76c5197cec636a0a3951afb36", "patch": "@@ -1105,7 +1105,7 @@ func DefaultOptions() Options {\n \t\tTopP: 0.9,\n \t\tTypicalP: 1.0,\n \t\tRepeatLastN: 64,\n-\t\tRepeatPenalty: 1.1,\n+\t\tRepeatPenalty: 1.0,\n \t\tPresencePenalty: 0.0,\n \t\tFrequencyPenalty: 0.0,\n \t\tSeed: -1," }, { "sha": "32b61935d34435c5b0b37d7e759d309cecf2dfc4", "filename": "docs/modelfile.mdx", "status": "modified", "additions": 1, "deletions": 1, "changes": 2, "blob_url": "https://github.com/ollama/ollama/blob/6a261db7d87b13d76c5197cec636a0a3951afb36/docs%2Fmodelfile.mdx", "raw_url": "https://github.com/ollama/ollama/raw/6a261db7d87b13d76c5197cec636a0a3951afb36/docs%2Fmodelfile.mdx", "contents_url": "https://api.github.com/repos/ollama/ollama/contents/docs%2Fmodelfile.mdx?ref=6a261db7d87b13d76c5197cec636a0a3951afb36", "patch": "@@ -148,7 +148,7 @@ PARAMETER \n | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |\n | num_ctx | Sets the size of the context window used to generate the next token. (Default: 2048) | int | num_ctx 4096 |\n | repeat_last_n | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx) | int | repeat_last_n 64 |\n-| repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1) | float | repeat_penalty 1.1 |\n+| repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.0, disabled) | float | repeat_penalty 1.1 |\n | temperature | The temperature of the model. Increasing the temperature will make the model answer more creatively. (Default: 0.8) | float | temperature 0.7 |\n | seed | Sets the random number seed to use for generation. Setting this to a specific number will make the model generate the same text for the same prompt. (Default: 0) | int | seed 42 |\n | stop | Sets the stop sequences to use. When this pattern is encountered the LLM will stop generating text and return. Multiple stop patterns may be set by specifying multiple separate `stop` parameters in a modelfile. | string | stop \"AI assistant:\" |" } ] }