{"slug": "tinker", "name": "Tinker", "category": "ai", "category_name": "AI tools and apps", "description": "", "recommenders": 3, "commenters": 9, "warners": 0, "score": 4.085, "rising": 0.35, "first": "2025-11-04", "latest": "2026-08-26", "why": "It\u2019s one of the smoother experiences we\u2019ve had running large scale RL workflows.", "why_is_quote": true, "has_verdict": false, "sentiment": {"recommend": 3, "neutral": 3, "compare": 3, "warn": 0}, "spark": [1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0], "listed": true, "loved": [], "caveat": "", "short_quote": "It\u2019s one of the smoother experiences we\u2019ve had running large scale RL workflows.", "kind": "", "icon": true, "domain": "thinkingmachines.ai", "github_repo": null, "official_url": "https://thinkingmachines.ai/tinker/", "pricing": "", "aliases": ["Tinker"], "verdict": null, "praise": [{"label": "", "summary": "", "quotes": [{"id": 791270433, "item_id": 49454402, "url": "https://news.ycombinator.com/item?id=49454402", "by": "quicheshore2003", "date": "2026-08-26", "version": null, "stance": "recommend", "quote": "Thanks to https://tinker-docs.thinkingmachines.ai/tinker/ for the training infra.", "excerpt": "Hi all, I'm Quiche. Here to share this cool project I've been working on. She's called colette. She's a persona model. She runs a restaurant named Tosca, grew up in Southern Italy, and is regarded as one of the best fine-dining chefs from that region. A persona model is an LM thats been given specific domain knowledge, a character background, and niche expertise via fine-tuning (SFT). No RL was used in this checkpoin", "context": "", "use_case": "Fine-tuning training infra", "switched_from": "", "story": "Show HN: Collete-01 - Person in a Model"}, {"id": 782882480, "item_id": 48930155, "url": "https://news.ycombinator.com/item?id=48930155", "by": "malshe", "date": "2026-07-15", "version": null, "stance": "recommend", "quote": "I used it last week for an application using a small model just as an experiment. It all went very well.", "excerpt": "They have good docs on finetuning in general here: https://thinkingmachines.ai/tinker/ I used it last week for an application using a small model just as an experiment. It all went very well. The model did not turn out to be good though because my training data was of bad quality. I plan to work on it more this weekend.", "context": "", "use_case": "fine-tuning a small model", "switched_from": "", "story": "Inkling: Our Open-Weights Model"}, {"id": 733034384, "item_id": 45814649, "url": "https://news.ycombinator.com/item?id=45814649", "by": "ramplabs", "date": "2025-11-04", "version": null, "stance": "recommend", "quote": "It\u2019s one of the smoother experiences we\u2019ve had running large scale RL workflows.", "excerpt": "We got early access to Tinker from Thinking Machines and spent the week putting it through its paces here at Ramp Labs. We used it to explore how RL post training performance changes when splitting data by domain and training an ensemble of specialized models versus a single model trained on everything at once. Tinker handled the heavy lifting like infrastructure, async rewards, and GPU orchestration so we could focu", "context": "on the Ramp Labs team", "use_case": "RL post-training workflows", "switched_from": "", "story": null}]}], "caveats": [], "who": [{"by": "Danau5tin", "context": "RL practitioner", "url": "https://news.ycombinator.com/item?id=48907077", "date": "2026-07-14"}, {"by": "ramplabs", "context": "on the Ramp Labs team", "url": "https://news.ycombinator.com/item?id=45814649", "date": "2025-11-04"}], "tags": [], "use_cases": ["fine-tuning training infra", "fine-tuning models with lora", "fine-tuning a small model", "finetuning infrastructure", "hosted fine-tuning", "rl training stack"], "compared": [{"slug": "claude", "name": "Claude", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [{"id": 790747728, "item_id": 49421733, "url": "https://news.ycombinator.com/item?id=49421733", "by": "alexgartrell", "date": "2026-08-24", "version": null, "stance": "recommend", "quote": "As a TML person, I'm obviously biased, but I couldn't resist because of", "excerpt": "https://thinkingmachines.ai/tinker/ https://github.com/thinking-machines-lab/tinker-cookbook (As a TML person, I'm obviously biased, but I couldn't resist because of \"tinkering\"). TBF it's hard to imagine a real architecture change that wouldn't require a ton of compute, but you could certainly fine tune and play with different recipes, loss functions, etc. And Claude can carry you a lot of the way through doing this", "context": "works at Thinking Machines (TML)", "use_case": "fine-tuning models", "switched_from": "", "story": "I were 17, I'd learn how to build LLMs from scratch"}], "threads": [{"url": "https://news.ycombinator.com/item?id=48924912", "title": "Inkling: Our Open-Weights Model", "mentions": 3}, {"url": "https://news.ycombinator.com/item?id=49454401", "title": "Show HN: Collete-01 - Person in a Model", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49412396", "title": "I were 17, I'd learn how to build LLMs from scratch", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49218466", "title": "Ask HN: How can I tell which LLM's have been trained on what?", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48905919", "title": "Show HN: I RL-trained an agent that trains models with RL (for \u2013$1.3k)", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48602963", "title": "I built a WhatsApp clone of myself; a friend agreed on a password with it", "mentions": 1}], "sentiment_by_month": {"2025-11": {"recommend": 1}, "2025-12": {}, "2026-01": {}, "2026-02": {}, "2026-03": {}, "2026-04": {}, "2026-05": {"compare": 1}, "2026-06": {"compare": 1}, "2026-07": {"neutral": 2, "compare": 1, "recommend": 1}, "2026-08": {"neutral": 1, "recommend": 1}, "2026-09": {}, "2026-10": {}}, "versions": [], "version_unnamed": 0, "family_ref": null, "runs_on": null, "lines": [], "quotes": [{"id": 791270433, "item_id": 49454402, "url": "https://news.ycombinator.com/item?id=49454402", "by": "quicheshore2003", "date": "2026-08-26", "version": null, "stance": "recommend", "quote": "Thanks to https://tinker-docs.thinkingmachines.ai/tinker/ for the training infra.", "excerpt": "Hi all, I'm Quiche. Here to share this cool project I've been working on. She's called colette. She's a persona model. She runs a restaurant named Tosca, grew up in Southern Italy, and is regarded as one of the best fine-dining chefs from that region. A persona model is an LM thats been given specific domain knowledge, a character background, and niche expertise via fine-tuning (SFT). No RL was used in this checkpoin", "context": "", "use_case": "Fine-tuning training infra", "switched_from": "", "story": "Show HN: Collete-01 - Person in a Model"}, {"id": 787503280, "item_id": 49218955, "url": "https://news.ycombinator.com/item?id=49218955", "by": "wmf", "date": "2026-08-07", "version": null, "stance": "neutral", "quote": "Some links I found but I can't vouch for them:", "excerpt": "Yeah, the dream is to have a \"reasoning kernel\" that you can plug specific knowledge and skills into (\"I know kung fu\"). I'm not aware of any progress on the dream though. Plenty of people are making LoRAs for general-purpose open models and there are companies like Thinking Machines that will help you. I'm not sure how you'd choose which base model to start with. I'm also not sure how to draw the boundary between Lo", "context": "", "use_case": "fine-tuning models with LoRA", "switched_from": "", "story": "Ask HN: How can I tell which LLM's have been trained on what?"}, {"id": 782882480, "item_id": 48930155, "url": "https://news.ycombinator.com/item?id=48930155", "by": "malshe", "date": "2026-07-15", "version": null, "stance": "recommend", "quote": "I used it last week for an application using a small model just as an experiment. It all went very well.", "excerpt": "They have good docs on finetuning in general here: https://thinkingmachines.ai/tinker/ I used it last week for an application using a small model just as an experiment. It all went very well. The model did not turn out to be good though because my training data was of bad quality. I plan to work on it more this weekend.", "context": "", "use_case": "fine-tuning a small model", "switched_from": "", "story": "Inkling: Our Open-Weights Model"}, {"id": 782878180, "item_id": 48929886, "url": "https://news.ycombinator.com/item?id=48929886", "by": "lebovic", "date": "2026-07-15", "version": null, "stance": "compare", "quote": "I found it easier and cheaper to just build from scratch.", "excerpt": "Aside from the raw capabilities, I'm excited to try Inkling for finetuning with the announced-but-not-yet-released small model. Ideally, I want to debug RL environments on a small model and then scale up to a large model. That's hard to do right now, in part because there's no good family of open-weight models with a great small and large MoE variant [1]. Concretely, my go-to small model for finetuning is Qwen 3.6 35", "context": "RL and finetuning practitioner", "use_case": "finetuning infrastructure", "switched_from": "", "story": "Inkling: Our Open-Weights Model"}, {"id": 782864768, "item_id": 48929048, "url": "https://news.ycombinator.com/item?id=48929048", "by": "tfehring", "date": "2026-07-15", "version": null, "stance": "neutral", "quote": "Thinky's main commercial product AFAIK is Tinker [0] - companies pay them to host their fine-tuning workloads", "excerpt": "Thinky's main commercial product AFAIK is Tinker [0] - companies pay them to host their fine-tuning workloads and then the resulting fine-tuned models. I don't know if this is a good business plan, but I'm sure at least one person there has read Joel on Software [1]. [0] https://thinkingmachines.ai/tinker/ [1] https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/", "context": "", "use_case": "hosted fine-tuning", "switched_from": "", "story": "Inkling: Our Open-Weights Model"}, {"id": 782513233, "item_id": 48907077, "url": "https://news.ycombinator.com/item?id=48907077", "by": "Danau5tin", "date": "2026-07-14", "version": null, "stance": "neutral", "quote": "the stack of Tinker, Prime-RL & Runpod - all of which I know well", "excerpt": "I chose the key technical decision and direction (such as the system architecture, the tasks to train on, the stack of Tinker, Prime-RL & Runpod - all of which I know well) etc. The problems it would do well on are training small agentic (multi-turn, tool use) task based models using the prime-rl stack, which are close to the distribution trained upon. It would likely not transfer to other training frameworks such as", "context": "RL practitioner", "use_case": "RL training stack", "switched_from": "", "story": "Show HN: I RL-trained an agent that trains models with RL (for \u2013$1.3k)"}, {"id": 777718720, "item_id": 48607420, "url": "https://news.ycombinator.com/item?id=48607420", "by": "dandinu", "date": "2026-06-20", "version": null, "stance": "compare", "quote": "even with a finetuning-as-a-serice like Tinker (the one from Thinking Machines) which is pretty cheap, the economics didn't work out that well.", "excerpt": "The method I used in the article is real and is pretty standard. Also, I do a decent amount of distillation and tinkering with weights for work, so I can assure you I did try that before resorting to good 'ol RAG. Overall, even with a finetuning-as-a-serice like Tinker (the one from Thinking Machines) which is pretty cheap, the economics didn't work out that well. Also, you probably one-shot this with Claude, I agree", "context": "does distillation and weight tinkering for work", "use_case": "fine-tuning a model", "switched_from": "", "story": "I built a WhatsApp clone of myself; a friend agreed on a password with it"}, {"id": 773352337, "item_id": 48334521, "url": "https://news.ycombinator.com/item?id=48334521", "by": "j316chuck", "date": "2026-05-30", "version": null, "stance": "compare", "quote": "Seems like an open source tinker", "excerpt": "Thoughts on trajectory and their multi-LoRA approach? Seems like an open source tinker but they have some cool results on harvey too https://trajectory.ai/field-notes/the-sovereign-path-for-continual-agent-learning-early-results-on-harvey-lab-with-nvidia-nemotron", "context": "", "use_case": "", "switched_from": "", "story": "Multi-Lora-Continual-Learning"}, {"id": 733034384, "item_id": 45814649, "url": "https://news.ycombinator.com/item?id=45814649", "by": "ramplabs", "date": "2025-11-04", "version": null, "stance": "recommend", "quote": "It\u2019s one of the smoother experiences we\u2019ve had running large scale RL workflows.", "excerpt": "We got early access to Tinker from Thinking Machines and spent the week putting it through its paces here at Ramp Labs. We used it to explore how RL post training performance changes when splitting data by domain and training an ensemble of specialized models versus a single model trained on everything at once. Tinker handled the heavy lifting like infrastructure, async rewards, and GPU orchestration so we could focu", "context": "on the Ramp Labs team", "use_case": "RL post-training workflows", "switched_from": "", "story": null}]}