--- title: Together AI expands fine-tuning service with more... description: Together AI has expanded its fine-tuning platform with support for the latest open-weight models (GLM 5.3, DeepSeek-V4-Flash, Kimi K2.7, Qwen 3.5/3.6/3.8,... canonical: https://daily.dev/posts/together-ai-expands-fine-tuning-service-with-more-models-live-metrics-and-finer-controls-jagreco8n twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Together AI expands fine-tuning service with more models, live metrics, and finer controls | daily.dev og:description: Together AI has expanded its fine-tuning platform with support for the latest open-weight models (GLM 5.3, DeepSeek-V4-Flash, Kimi K2.7, Qwen 3.5/3.6/3.8,... og:url: https://daily.dev/posts/together-ai-expands-fine-tuning-service-with-more-models-live-metrics-and-finer-controls-jagreco8n og:image: https://api.daily.dev/og/posts/JAGRECO8n.png og:image:alt: Together AI expands fine-tuning service with more models, live metrics, and finer controls og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Together AI expands fine-tuning service with more models, live metrics, and finer controls **[Together AI](https://daily.dev/sources/togetherai)** · 10 min read · 0 upvotes · 0 comments ## Summary Together AI has expanded its fine-tuning platform with support for the latest open-weight models (GLM 5.3, DeepSeek-V4-Flash, Kimi K2.7, Qwen 3.5/3.6/3.8, Gemma 4), live experiment tracking via API/CLI/dashboard, and new training controls including Expert LoRA (which trains MoE expert layers rather than just attention, recalling up to 89% of new knowledge vs 15% for attention-only adapters), early stopping with automatic refunds for unused training steps, arbitrary batch sizes via gradient accumulation, tokenized dataset previews, per-example sample weights, pre-flight file validation, and configurable sequence packing. Training prices dropped 30-70% on select models, with gpt-oss-20b SFT falling from $1.50 to $0.40 per million tokens. A case study covers Adaption's AutoScientist, which runs closed-loop self-improving training up to 1 trillion parameters on the platform. Upcoming: deploying intermediate LoRA adapters while training is still running, starting with GLM-5.3 and Kimi K3. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is Expert LoRA and how does it improve fine-tuning of Mixture-of-Experts models? Expert LoRA is a technique that attaches LoRA adapters directly to the expert layers of a Mixture-of-Experts model instead of only attention layers, since over 90% of an MoE model's parameters and most of its knowledge live in the experts. In tests teaching models 200 invented facts, expert adapters recalled up to 89% of new knowledge versus 15% for attention-only adapters, and scored 75.3% vs 71.5% on MMLU-Pro. Enable it by adding expert modules to lora_trainable_modules. _Teams weighing LoRA strategies for MoE models can track platform-level fine-tuning changes like this on daily.dev._ ### How much did Together AI cut fine-tuning training prices for gpt-oss models? Training prices for gpt-oss-20b SFT dropped from $1.50 to $0.40 per million tokens, roughly a 70% reduction, while gpt-oss-120b SFT fell from $5.00 to $2.50 per million tokens. DPO training for gpt-oss-20b dropped from $3.75 to $1.00 per million tokens. Price cuts of 30% to 70% were applied across most supported models, including Qwen3.5-9B, Qwen3.8-27B, and Gemma-4-31B. _Anyone budgeting fine-tuning costs across providers can follow pricing shifts like these on daily.dev._ ### How does early stopping work in Together AI's fine-tuning service? Early stopping automatically halts a fine-tuning run once validation loss plateaus, keeping the checkpoint with the best validation loss as the final model rather than the last one produced. It requires a validation set, watches validation loss at every evaluation step, and automatically refunds every training step that wasn't used. Patience, sensitivity to small improvements, and warmup are all tunable parameters. _Developers tuning training runs to avoid wasted compute can keep up with features like this via daily.dev._ ## Similar posts on daily.dev - [Together Post-Training: frontier-grade model shaping, the way that fits you](https://daily.dev/posts/together-post-training-frontier-grade-model-shaping-the-way-that-fits-you-8tnbnhhms) · Together AI · 1 upvotes · 0 comments - [Mastering Model Adaptation: A Guide to Fine-Tuning on Google Cloud](https://daily.dev/posts/mastering-model-adaptation-a-guide-to-fine-tuning-on-google-cloud-lg6pheaxe) · Google Cloud · 1 upvotes · 0 comments - [Claude, ChatGPT, and Gemini get all the hype, but the most interesting AI models are coming from elsewhere](https://daily.dev/posts/claude-chatgpt-and-gemini-get-all-the-hype-but-the-most-interesting-ai-models-are-coming-from-els-glbavr35d) · XDA Developers · 0 upvotes · 0 comments - [Autonomous LLM post-training with Tunix on TPUs](https://daily.dev/posts/autonomous-llm-post-training-with-tunix-on-tpus-fwvhv3x0a) · Google Developers · 1 upvotes · 0 comments --- Tags: [#deep-learning](https://daily.dev/tags/deep-learning), [#llm](https://daily.dev/tags/llm), [#lora](https://daily.dev/tags/lora), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts), [#together-ai](https://daily.dev/tags/together-ai) [View this post on daily.dev](https://daily.dev/posts/together-ai-expands-fine-tuning-service-with-more-models-live-metrics-and-finer-controls-jagreco8n) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Together AI expands fine-tuning service with more models, live metrics, and finer controls","url":"https://daily.dev/posts/together-ai-expands-fine-tuning-service-with-more-models-live-metrics-and-finer-controls-jagreco8n","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/together-ai-expands-fine-tuning-service-with-more-models-live-metrics-and-finer-controls-jagreco8n"},"datePublished":"2026-09-11T15:57:19.790Z","dateModified":"2026-10-15T02:22:42.821Z","description":"Together AI has expanded its fine-tuning platform with support for the latest open-weight models (GLM 5.3, DeepSeek-V4-Flash, Kimi K2.7, Qwen 3.5/3.6/3.8,...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/e1fd35361980deb3abdd3a6870da238b?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/e1fd35361980deb3abdd3a6870da238b?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Together AI","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Together AI","logo":"https://media.daily.dev/image/upload/s--tCjWcJfJ--/f_auto,q_auto/v1780213200/logos/togetherai?_a=BAMAMiWQ0","url":"https://daily.dev/sources/togetherai"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/together-ai-expands-fine-tuning-service-with-more-models-live-metrics-and-finer-controls-jagreco8n","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"deep-learning,llm,lora,mixture-of-experts,together-ai","timeRequired":"PT10M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Together AI","item":"https://daily.dev/sources/togetherai"},{"@type":"ListItem","position":3,"name":"Together AI expands fine-tuning service with more models, live metrics, and finer controls"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/together-ai-expands-fine-tuning-service-with-more-models-live-metrics-and-finer-controls-jagreco8n#faq","mainEntity":[{"@type":"Question","name":"What is Expert LoRA and how does it improve fine-tuning of Mixture-of-Experts models?","acceptedAnswer":{"@type":"Answer","text":"Expert LoRA is a technique that attaches LoRA adapters directly to the expert layers of a Mixture-of-Experts model instead of only attention layers, since over 90% of an MoE model's parameters and most of its knowledge live in the experts. In tests teaching models 200 invented facts, expert adapters recalled up to 89% of new knowledge versus 15% for attention-only adapters, and scored 75.3% vs 71.5% on MMLU-Pro. Enable it by adding expert modules to lora_trainable_modules. Teams weighing LoRA strategies for MoE models can track platform-level fine-tuning changes like this on daily.dev."}},{"@type":"Question","name":"How much did Together AI cut fine-tuning training prices for gpt-oss models?","acceptedAnswer":{"@type":"Answer","text":"Training prices for gpt-oss-20b SFT dropped from $1.50 to $0.40 per million tokens, roughly a 70% reduction, while gpt-oss-120b SFT fell from $5.00 to $2.50 per million tokens. DPO training for gpt-oss-20b dropped from $3.75 to $1.00 per million tokens. Price cuts of 30% to 70% were applied across most supported models, including Qwen3.5-9B, Qwen3.8-27B, and Gemma-4-31B. Anyone budgeting fine-tuning costs across providers can follow pricing shifts like these on daily.dev."}},{"@type":"Question","name":"How does early stopping work in Together AI's fine-tuning service?","acceptedAnswer":{"@type":"Answer","text":"Early stopping automatically halts a fine-tuning run once validation loss plateaus, keeping the checkpoint with the best validation loss as the final model rather than the last one produced. It requires a validation set, watches validation loss at every evaluation step, and automatically refunds every training step that wasn't used. Patience, sensitivity to small improvements, and warmup are all tunable parameters. Developers tuning training runs to avoid wasted compute can keep up with features like this via daily.dev."}}]} ```