Changelog - Together AI docs
Documentation Index Fetch the complete documentation index at: /llms.txt Use this file to discover all available pages before exploring further. Skip to main content Assistant Responses are generated using AI and may contain mistakes. New models New serverless models The following models are now available on serverless : tencent/Hy4 : 1,048,576 context length, NVFP4 quantization, text and image input, reasoning, function calling, and structured outputs.
Pricing: $0.834 input / $2.501 output / $0.042 cached input (per 1M tokens). New releases Together Python SDK and CLI 3.3.0 What’s new: Tinker 0.39.0 compatibility: The Tinker-compatible client now builds on tinker 0.39.0, which the tinker extra installs. It follows Tinker 0.39’s call styles, so inside a running event loop you can await the same calls you make in synchronous code. ServiceClient now works as a context manager that closes every training client it created. forward_backward now returns per-datum log-probabilities by default, as Tinker does.
Deployment placement in Jig: The [tool.jig.deploy] section of your Jig config accepts two placement settings. placement_group puts every deployment that shares its value on one cluster, and regions limits a deployment to clusters in the regions you list. tg beta jig status shows both when they’re set. New releases New models New serverless models The following models are now available on serverless : Qwen/Qwen4-Max : 1,000,000 context length, text and image input, reasoning, function calling, and structured outputs.
Pricing: $1.425 input / $4.275 output / $0.07125 cached input (per 1M tokens). New models available for fine-tuning You can now fine-tune the following models: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B Both models support function-calling and reasoning fine-tuning. See Supported models for the full list. Together Python SDK and CLI 3.2.1 What’s new: Forced cleanup of model resources: A ModelResourcesClient handle that created its resource now stops it with force=True when its context manager exits, which also stops every session still attached.
The cleanup after a failed create forces the stop too, and raises the stop’s error if the stop fails. stop returns sooner: ModelResourcesClient.stop now returns as soon as the API accepts the stop , which is when billing ends. It no longer waits for the resource to finish stopping. Evaluations removed from the CLI: Following the Evaluations shutdown, tg evals commands fail with Service discontinued , and tg files upload no longer accepts --purpose eval .
The client.evals methods are marked as deprecated. New releases Improvements Together Python SDK and CLI 3.2.0 What’s new: Rollout capacity budgets: tg endpoints rollout takes two flags that set a rollout’s capacity budgets . --max-surge caps how many replicas each canary step adds to the target. --max-unavailable lets a rollout start from a DEGRADED source, or shut down source replicas to make room for the target.
Config revision lookup: tg models configs get returns a config revision by its ID ( cr_... ). New shared volumes in scripted cluster creates: tg clusters create takes --shared-volume-name and --shared-volume-size-tib , which create a shared volume with the cluster. Before this release, a --non-interactive create could only attach an existing volume. Attached model resources keep running: An attached ModelResourcesClient handle used as a context manager now detaches on exit and leaves the resource running.
Before this release, it stopped the resource. Default expiry for Jig secrets New Jig secrets now expire after 14 days unless you set a shorter lifetime with --ttl . Before a secret expires, the project’s admins and the secret’s creator get a daily email reminder. An expired secret is deleted, so new replicas can’t read it. Existing replicas keep the value until they terminate. To renew a secret, set a new value or set it again with --ttl .
New models Improvements New dedicated endpoint models The following models are now available for deployment on dedicated endpoints : Qwen/Qwen3-TTS-12Hz-1.7B (text-to-speech). This model was previously listed as Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice . thinkingmachines/Inkling-2-Small . Post-training: Nemotron 3.5 Super VL nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 is available for post-training : low-rank adapter (LoRA) training on B200-SXM , 163,840-token training and generator contexts, rank up to 32, and up to three adapter sessions per resource.
The cookbook has no codec for it yet, so this model is currently SDK-only. Deployment support is coming soon. See Supported models for more details. Post-training: sampling-only model resources A post-training model resource can now run generators with no trainer. To sample a saved inference checkpoint without training it further, set num_trainer_replicas=0 . You’re billed only for the generator GPUs. See Sampling-only resources .
New models Deprecations New serverless models The following models are now available on serverless : thinkingmachines/Inkling-2-Small : 1,048,576 context length, NVFP4 quantization, text and image input, function calling, and structured outputs. Pricing: $0.30 input / $1.00 output / $0.03 cached input (per 1M tokens). Qwen/Qwen3-TTS-12Hz-1.7B : text-to-speech with nine preset voices in 10 languages. Pricing: $20.00 per 1M characters.
Evaluations retired Evaluations have been shut down and are no longer available. This covers the evaluations API, the client.evals methods in the Python and TypeScript SDKs, the tg evals CLI commands, and the Evaluations page in the console. For questions, contact support . New models New dedicated endpoint models The following models are now available for deployment on dedicated endpoints : nvidia/nemotron-3-asr-streaming-0.6b (speech-to-text). nvidia/nemotron-3.5-asr-streaming-0.6b (speech-to-text).
New models Improvements Deprecations New serverless models The following models are now available on serverless : elorian-labs/glimpse-1.0 : 262,144 context length, FP8 quantization, function calling, and structured outputs. Pricing: $0.60 input / $3.00 output / $0.06 cached input (per 1M tokens). New provisioned throughput models The following models are now available on provisioned throughput : moonshotai/Kimi-K3.1-Pro .
Rate metrics for dedicated endpoints The metrics endpoint for dedicated endpoints now exports request, token, and latency metrics ending in _rate5m . They’re precomputed per-second rates over the last five minutes, so you can chart them without rate() . They don’t report false spikes when a replica restarts, scales down, or goes idle. The counters and histograms they replace are deprecated, and Together AI plans to remove them.
See Deprecated metrics for replacements and query migration. New models New dedicated endpoint models The following models are now available for deployment on dedicated endpoints : ricursive-ai/Ripple-65B-Preview . Improvements Source validation for remote model uploads Remote uploads of custom models and adapters now reject unsupported sources, such as s3:// paths and http:// URLs, before the upload starts. The source must be a Hugging Face repo ID or URL, or a presigned https:// URL on Amazon S3, Google Cloud Storage, or Cloudflare R2.
To import from an s3:// path, pass a presigned URL for the object instead. New models Improvements New serverless models The following models are now available on serverless : moonshotai/Kimi-K3.1-Pro : 1,048,576 context length, NVFP4 quantization, text and image input, function calling, and structured outputs. Pricing: $2.00 input / $10.00 output / $0.20 cached input (per 1M tokens). ricursive-ai/Ripple-65B-Preview : 131,072 context length, NVFP4 quantization, function calling, and structured outputs.
Pricing: $0.20 input / $0.60 output / $0.02 cached input (per 1M tokens). Messages API and Responses API support more models The Messages API and R
The rest of this page is on the live site.