Heads up This site is currently under heavy development.
← all tools
◆ AI Model & Data Infrastructure

Together AI

changelog-20260831-07a75e81 commercial

Run, train, and serve open-source AI models on Together AI.

source ↗ rss ↗

Summary

Run, train, and serve open-source AI models on Together AI.

Release history

  1. changelog-20260831-07a75e81 Aug 31, 2026 · issue 012
    └──▷ HOW TO FIND IT
    Create a project API key with a 7-day expiration to issue short-lived credentials for a contractor or CI pipeline.
    📍In the console, go to Settings › API Keys, click 'Create API key', select 'Set an expiration date', choose '7 days', and save.
    • ›Adds tg batches command group to the Together CLI with subcommands tg batches submit (upload a local JSONL file or pass an existing file ID to create a batch job), tg batches list, tg batches get (alias ls), tg batches cancel, and tg batches download (streams results to stdout or writes output/error files to disk with --output).
    • ›API key expiration is now configurable at creation time in the console: choose 1 hour, 1 day, 7 days, 30 days, or a custom date via Set an expiration date.
    • ›Connecting a Vercel project via Integrations settings now creates a dedicated API key per linked Vercel project, automatically set as the TOGETHER_API_KEY environment variable in that Vercel project.
    • ›Project collaborators can now leave a project themselves from Settings > Project > Collaborators or the Projects list (organization members cannot leave the default project; last admin must promote another before leaving).
  2. changelog-20260830-07a75e81 Aug 30, 2026 · issue 011

    Together AI adds a tg batches CLI command group, API key expiration controls, ACH payments, and new GLM-5.3 models with 1M-token context.

    └──▷ TRY IT
    Submit a local JSONL file as a batch inference job and then download results to disk once complete.
    $ tg batches submit requests.jsonl
tg batches list
tg batches download <job-id> --output results/
    • ›Adds tg batches command group to the Together CLI with subcommands tg batches submit (upload a JSONL file or pass an existing file ID to create a batch job), tg batches list, tg batches cancel, tg batches download (stream results to stdout or write files to disk with --output); ls and get are supported as aliases.
    • ›Fine-tuning training quality improved for Qwen/Qwen3.5-0.8B, Qwen3.5-2B, Qwen3.5-4B, Qwen3.5-27B, Qwen3.5-35B-A3B, Qwen3.5-35B-A3B-Base, Qwen3.5-122B-A10B, nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16, and nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 — start a new job with existing data and settings to pick up the improvement.
  3. changelog-20260829-07a75e81 Aug 29, 2026 · issue 010

    Together CLI gains tg batches command group; new GLM-5.3 models, API key expiration, multi-project GA, and ACH payments launch

    └──▷ TRY IT
    Submit a JSONL batch inference job from the CLI and download results once complete — useful for high-volume, cost-efficient offline inference.
    $ tg batches submit requests.jsonl
tg batches list
tg batches download <batch-id> --output results/
    • ›Adds tg batches CLI command group for batch inference, with subcommands tg batches submit (upload a local JSONL file or existing file ID), tg batches list, tg batches cancel, tg batches download (streams to stdout or writes to disk with --output), and aliases ls and get for job lifecycle management.
    • ›Adds new model zai-org/GLM-5.3 with 1,000,000-token context, FP4 quantization, function calling, and structured outputs; priced at $1.40 input / $4.40 output / $0.26 cached input per 1M tokens.
    • ›Adds new model zai-org/GLM-5.3-Flash with 1,000,000-token context, FP8 quantization, function calling, and structured outputs; priced at $0.15 input / $0.50 output / $0.03 cached input per 1M tokens.
    • ›Adds Cancel expiration action (via the three-dot menu) to remove a scheduled expiration from an existing project API key.
    • ›Multi-project isolation is now generally available for all organizations — clusters, fine-tuned models, endpoints, evaluations, files, and API keys are fully scoped to projects with no early-access limitations.
    +2 moreshow less
    • ›ACH bank transfers are now available to all customers (previously enterprise-only); link a U.S. bank account via billing settings with instant verification and purchase credits directly.
    • ›Fine-tuning quality improved for Qwen/Qwen3.5-0.8B, Qwen/Qwen3.5-2B, Qwen/Qwen3.5-4B, Qwen/Qwen3.5-27B, Qwen/Qwen3.5-35B-A3B, Qwen/Qwen3.5-35B-A3B-Base, Qwen/Qwen3.5-122B-A10B, nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16, and nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 — no setting changes required; start a new job to pick it up.
  4. 2.0.0 Aug 28, 2026 · issue 009

    API surface changed: +2 endpoints, 8 modified

    API CHANGE

    API surface changed: +2 endpoints, 8 modified

    • + GET /billing/usage
    • + GET /rl/checkpoints/{id}
    • ~ GET /rl/model-resources/{model_resources_id}: response schema changed
    • ~ POST /projects/{projectId}/endpoints/{endpointId}/rollouts: request body changed
    • ~ POST /projects/{projectId}/endpoints/{endpointId}/rollouts/preview-defaults: request body changed
    • ~ POST /rl/model-resources: response schema changed
    • ~ POST /rl/model-resources/{model_resources_id}/stop: response schema changed
    • ~ POST /rl/training-sessions/{session_id}/operations/custom-forward-backward: request body changed
    • ~ POST /rl/training-sessions/{session_id}/operations/forward: request body changed
    • ~ POST /rl/training-sessions/{session_id}/operations/forward-backward: request body changed
    • ›New endpoint GET /billing/usage
    • ›New endpoint GET /rl/checkpoints/{id}
    • ›GET /rl/model-resources/{model_resources_id}: response schema changed
    • ›POST /projects/{projectId}/endpoints/{endpointId}/rollouts: request body changed
    • ›POST /projects/{projectId}/endpoints/{endpointId}/rollouts/preview-defaults: request body changed
    +5 moreshow less
    • ›POST /rl/model-resources: response schema changed
    • ›POST /rl/model-resources/{model_resources_id}/stop: response schema changed
    • ›POST /rl/training-sessions/{session_id}/operations/custom-forward-backward: request body changed
    • ›POST /rl/training-sessions/{session_id}/operations/forward: request body changed
    • ›POST /rl/training-sessions/{session_id}/operations/forward-backward: request body changed
  5. docs update Aug 28, 2026 · issue 009

    Together AI fine-tuning file uploads now support a progress_callback via FileUploadProgress events in the Python SDK.

    • ›Adds optional progress_callback parameter to file uploads, accepting a callable that receives FileUploadProgress events exposing event.uploaded_bytes and event.total_bytes for real-time upload progress tracking.
  6. docs update Aug 28, 2026 · issue 009

    Together AI GPU clusters now expose kubeconfig download via console, CLI get-credentials, and API for both Kubernetes and Slurm clusters.

    └──▷ TRY IT
    Download and set the default kubeconfig for a specific GPU cluster so kubectl commands target it immediately.
    $ tg beta clusters get-credentials <CLUSTER_ID> --set-default-context
    Look up the cluster ID needed before running get-credentials.
    $ tg beta clusters list
    Download the kubeconfig from the console when OIDC is enabled and you need the admin kubeconfig for RBAC setup.
    📍1. Go to api.together.ai/clusters and open your cluster. 2. In the cluster sidebar, find the 'Admin Kubeconfig' row. 3. Click 'View' to copy the contents, or 'Download' to save it as a file.
    • ›Adds tg beta clusters get-credentials [CLUSTER_ID] --set-default-context CLI command to download and set the kubeconfig for a GPU cluster.
    • ›Adds tg beta clusters list CLI command to enumerate cluster IDs.
    • ›Exposes kubeconfig download (view/copy or save as file) from the cluster details page at api.together.ai/clusters once a cluster is ready.
    • ›Introduces role-scoped kubeconfig access: the Kubeconfig row is visible to all project members; when OIDC is enabled, the Admin Kubeconfig row is restricted to project admins only, while the OIDC Kubeconfig row remains accessible to all project members (admins and editors).
    • ›Both Kubernetes and Slurm cluster types expose a kubeconfig for access to the cluster's Kubernetes API.
  7. docs update Aug 28, 2026 · issue 009
    └──▷ HOW TO FIND IT
    Access the kubeconfig from the console when you need to inspect cluster resources or run kubectl commands against your Slurm cluster.
    📍1. Open the Together AI console and navigate to your GPU cluster. 2. Open the cluster details page. 3. Click 'Download kubeconfig' to retrieve the file.
    • ›Enables any project member to download the kubeconfig for a Slurm-backed GPU cluster via the Together CLI or from the cluster details page, granting direct access to the underlying Kubernetes API.
  8. docs update Aug 28, 2026 · issue 009

    Together AI adds a tg batches CLI command group, ACH bank transfers, per-key expiration, and a 1M-context GLM-5.3-Flash model.

    └──▷ TRY IT
    Download completed batch results to disk once the job finishes, capturing both output and error files.
    $ tg batches download <job-id> --output ./results/
    Set a 30-day expiration on a new API key to enforce key rotation policy for a project.
    📍In the console, go to Settings › Project › API Keys, click 'Create key', select 'Set an expiration date', then choose '30 days', and confirm.
    • ›New tg batches command group in the Together CLI enables batch inference: tg batches submit uploads a local JSONL file or existing file ID and creates a job; tg batches list, get (alias ls), cancel, and download manage the job lifecycle; tg batches download streams results to stdout or writes output and error files to disk with --output.
    • ›New model zai-org/GLM-5.3-Flash available on serverless: 1,000,000-token context length, FP8 quantization, function calling and structured outputs, priced at $0.15 input / $0.50 output / $0.03 cached input per 1M tokens.
    • ›ACH bank transfers now generally available to all customers — link a U.S. bank account with instant verification from billing settings and purchase credits directly; credits deposit after 1–3 business days.
    • ›Multi-project isolation is now enabled for every organization, with clusters, fine-tuned models, endpoints, evaluations, files, and API keys fully scoped to projects; early-access limitations removed.
    • ›API key expiration can now be set at creation time in the console — choose 1 hour, 1 day, 7 days, 30 days, or a custom date.
    +2 moreshow less
    • ›Scheduled expiration on a project API key can now be cancelled in the console via the three-dot menu (Cancel expiration), keeping the key active indefinitely.
    • ›Project collaborators can now leave a project themselves from Settings > Project > Collaborators or the Projects list, without requiring an admin action.
  9. docs update Aug 28, 2026 · issue 009
    └──▷ TRY IT
    Submit a JSONL file of chat completion requests as a batch job to process them asynchronously at scale.
    $ tg batches submit ./requests.jsonl chat.completions
    Download the output of a completed batch job to a local file for downstream processing.
    $ tg batches get [BATCH_ID] --output ./out
    List all batch jobs and cancel one that is no longer needed.
    $ tg batches ls
tg batches cancel [BATCH_ID]
    • ›Adds batches subcommand to the tg CLI with submit, get, cancel, and ls operations for managing batch inference jobs.
    • ›New tg batches submit ./requests.jsonl chat.completions command submits a JSONL file of requests as a batch inference job.
    • ›New tg batches get [BATCH_ID] --output ./out retrieves a completed batch job and writes results to a local path.
    • ›New tg batches cancel [BATCH_ID] cancels an in-progress batch job.
    • ›New tg batches ls lists all batch inference jobs.
  10. docs update Aug 28, 2026 · issue 009

    Together AI adds dedicated endpoints with reserved hardware, LoRA adapter serving, Seedance 2.5 video generation, and a full billing/endpoint management API.

    • ›New POST /v1/endpoints API (createendpoint) deploys a dedicated, reserved-hardware inference endpoint for any supported model — starts automatically after creation.
    • ›New GET /v1/endpoints API (listendpoints) lists all endpoints for your account, filterable by type (dedicated or serverless).
    • ›New GET /v1/endpoints/{id} API (getendpoint) retrieves current state, configuration, and scaling settings for a specific endpoint.
    • ›New PATCH /v1/endpoints/{id} API (updateendpoint) updates display name, autoscaling settings, or endpoint state (start/stop) on an existing dedicated endpoint.
    • ›New DELETE /v1/endpoints/{id} API (deleteendpoint) permanently removes a dedicated endpoint.
    +10 moreshow less
    • ›New POST /v1/endpoints/{id}/adapters API (addadapter) attaches a LoRA adapter to a running dedicated endpoint, routing inference requests by adapter model name.
    • ›New GET /v1/endpoints/{id}/adapters API (listadapters) returns all LoRA adapters bound to a specified dedicated endpoint.
    • ›New DELETE /v1/endpoints/{id}/adapters/{adapter} API (removeadapter) unbinds a LoRA adapter from a dedicated endpoint.
    • ›New GET /v1/hardware API (listhardware) lists available hardware configurations for model deployment; accepts a model parameter to filter to compatible hardware and show availability status.
    • ›New GET /v1/billing/usage API (billing-usage) returns an organization's monthly billing usage as cost-annotated line items grouped by time window, at daily or hourly granularity (UTC).
    • ›New POST /v1/models/upload API (upload-model) uploads a custom or fine-tuned model from Hugging Face or S3 for serving on a dedicated endpoint.
    • ›Dedicated endpoints support per-endpoint configuration of replica count, hardware, decoding optimizations, and prompt caching via endpoint settings.
    • ›Supports serving multiple LoRA adapters simultaneously on a single dedicated endpoint, with attach/list/remove lifecycle management.
    • ›New batches CLI subcommand lets practitioners submit, monitor, download, and cancel batch inference jobs from the terminal.
    • ›New Seedance 2.5 model support enables multi-shot video generation with synchronized audio from text, image, video, and audio inputs.
  11. Batch jobs in the CLI Aug 26, 2026 · issue 009

    Together CLI gains tg batches command group for submitting and managing batch inference jobs.

    └──▷ TRY IT
    List all batch jobs and cancel one that is no longer needed.
    $ tg batches ls
tg batches cancel <job-id>
    • ›Adds tg batches submit to upload a local JSONL file (or reference an existing file ID) and create a batch inference job against chat.completions, audio.transcriptions, or audio.translations.
    • ›Adds tg batches list (alias ls), tg batches retrieve (alias get), and tg batches cancel to manage the full batch job lifecycle.
    • ›Adds tg batches download to stream results to stdout or write output and error files to disk via --output.
  12. 2.0.0 Aug 24, 2026 · issue 006

    API surface changed: 8 modified

    API CHANGE

    API surface changed: 8 modified

    • ~ GET /compute/clusters/{cluster_id}/instances/{instance_id}/remediations/{remediation_id}: response schema changed
    • ~ GET /compute/passive_health_check_alerts: response schema changed
    • ~ POST /compute/clusters/{cluster_id}/instances/{instance_id}/remediations: request body changed; response schema changed
    • ~ POST /compute/clusters/{cluster_id}/instances/{instance_id}/remediations/{remediation_id}/approve: response schema changed
    • ~ POST /compute/clusters/{cluster_id}/instances/{instance_id}/remediations/{remediation_id}/cancel: response schema changed
    • ~ POST /compute/clusters/{cluster_id}/instances/{instance_id}/remediations/{remediation_id}/reject: response schema changed
    • ~ POST /rl/training-sessions: request body changed
    • ~ POST /rl/training-sessions/{session_id}/operations/sample: request body changed
    • ›GET /compute/clusters/{cluster_id}/instances/{instance_id}/remediations/{remediation_id}: response schema changed
    • ›GET /compute/passive_health_check_alerts: response schema changed
    • ›POST /compute/clusters/{cluster_id}/instances/{instance_id}/remediations: request body changed; response schema changed
    • ›POST /compute/clusters/{cluster_id}/instances/{instance_id}/remediations/{remediation_id}/approve: response schema changed
    • ›POST /compute/clusters/{cluster_id}/instances/{instance_id}/remediations/{remediation_id}/cancel: response schema changed
    +3 moreshow less
    • ›POST /compute/clusters/{cluster_id}/instances/{instance_id}/remediations/{remediation_id}/reject: response schema changed
    • ›POST /rl/training-sessions: request body changed
    • ›POST /rl/training-sessions/{session_id}/operations/sample: request body changed
  13. 2.0.0 Aug 20, 2026 · issue 002

    Together AI now publishes an API — 197 endpoints across 42 areas: Rl, Fine Tuning, Volumeservice, …

    • ›Rl (28 endpoints) — create, read, update
    • ›Fine Tuning (14 endpoints) — create, read, delete
    • ›Volumeservice (13 endpoints) — create, read, update, delete
    • ›Deploymentservice (10 endpoints) — create, read, update, delete
    • ›Rolloutservice (10 endpoints) — create, read, delete
    +4 moreshow less
    • ›Shadowexperimentservice (10 endpoints) — create, read, update, delete
    • ›Endpointservice (9 endpoints) — create, read, update, delete
    • ›Endpoints (8 endpoints) — create, read, update, delete
    • ›34 more areas: Deployments, Remediationservice, Abexperimentservice, Audio, Files, Gpuclusterservice, Instanceclusteraddonservice, Queue, Secrets, Sharedvolumeservice, Evaluation, Batches, Volumes, Code Interpreter, Configservice, Deploymentsstorage, Deploymentsvolumes, Instancetypeservice, Jobs, Models, Placementprofileservice, Supportedmodelsservice, Video, Account, Chat, Completion, Embeddings, Hardware, Healthcheckservice, Images, Regionservice, Rerank, Voices, Endpoints
  14. ACH bank transfers generally available Aug 19, 2026 · issue 009

    Together AI opens ACH bank transfers to all customers for direct credit purchases from U.S. bank accounts.

    • ›Supports ACH bank transfers as a payment method for all customers (previously enterprise-only): link a U.S. bank account with instant verification in billing settings, set it as default, and purchase credits directly — credits deposit after ACH clears (typically 1–3 business days).
  15. Fully automatic confirmation policy for node auto repair Aug 18, 2026 · issue 003

    Together AI adds fully automatic node repair policy with per-fault scoping and job interruption controls

    └──▷ HOW TO FIND IT
    Enable fully automatic repair for transient faults while keeping destructive actions gated, protecting running jobs with a grace period.
    📍In the console, go to the Repairs tab › Auto-remediation policy › set Confirmation policy to 'Fully automatic' › open Repair actions and enable 'VM reboot' for unattended execution, leave 'Reprovision' and 'Migrate to new host' requiring approval › set Wait policy to 'Grace period' to protect in-flight jobs.
    • ›New Fully automatic confirmation policy option under Auto-remediation policy on the Repairs tab lets auto node repair execute end-to-end without manual approval steps.
    • ›Per-fault scoping via Repair actions lets operators select which fault groups — Migrate to new host, Reprovision, VM reboot — run unattended under Fully automatic, keeping destructive repairs gated while transient ones self-clear.
  16. Project visibility Aug 17, 2026 · issue 002

    Together AI projects gain three visibility levels and full multi-resource scoping across clusters, models, endpoints, and API keys.

    └──▷ HOW TO FIND IT
    Change an existing project's visibility to Private so only collaborators and org admins can see it.
    📍In the console, go to your Project › Project Settings, then set the Visibility field to 'Private' and save.
    • ›Adds three project visibility levels — Open (any org member can join), Closed (discoverable but join requires admin approval), and Private (visible only to existing collaborators and org admins) — configurable at creation time or later via Project Settings.
    • ›Clusters, fine-tuned models, endpoints, evaluations, files, and API keys are now fully scoped to projects, removing early-access limitations on project isolation.
  17. New models available for fine-tuning Aug 13, 2026 · issue 002

    Together AI adds fine-tuning model limits CLI, tokenized dataset download, endpoint events, and multi-project resource scoping

    └──▷ TRY IT
    Check a model's fine-tuning constraints before launching a job, to validate your hyperparameter choices against platform limits.
    $ tg ft model-limits zai-org/GLM-5.2 --json
    After a fine-tuning job completes, download the tokenized dataset to audit the exact token sequences the model was trained on.
    $ tg ft download-tokenized-dataset <JOB_ID>
    Monitor a dedicated endpoint's lifecycle in real time — replica scaling, traffic shifts, and status changes — without leaving the terminal.
    $ tg beta endpoints events
    • ›Adds tg fine-tuning model-limits (alias tg ft model-limits <model>) subcommand, which queries GET /fine-tunes/models/limits to return a base model's capability flags and hyperparameter bounds including sequence-length, batch-size, and LoRA rank limits; supports --json for the full response body.
    • ›Adds tg ft download-tokenized-dataset <ft_id> subcommand to download the tokenized dataset archive a fine-tuning job trained on, enabling audit of exactly what the model saw; the CLI command also surfaces in the tokenized_dataset_upload_complete event message.
    • ›The tokenized_dataset_upload_complete event message — visible in the console Events tab and in GET /fine-tunes/{id}/events — now includes the ready-to-run tg ft download-tokenized-dataset <JOB_ID> command when a fine-tuning job finishes uploading its tokenized dataset archive.
    • ›Adds tg beta endpoints events subcommand to list a dedicated endpoint's audit and lifecycle events from the terminal, including replica scaling, traffic shifts, status changes, and pauses across every deployment under the endpoint.
    • ›Projects now support three visibility levels — Open, Closed, and Private — configurable at creation time or changed later from Project Settings.
    +3 moreshow less
    • ›Clusters, fine-tuned models, endpoints, evaluations, files, and API keys are now fully scoped to projects, removing early-access limitations on project isolation.
    • ›Adds zai-org/GLM-5.2 as a new supported model for fine-tuning.
    • ›Adds deepseek-ai/DeepSeek-V4-Flash-0731 as a new model available for deployment on dedicated endpoints.
  18. Leave a project from the console Aug 12, 2026 · issue 009

    Together AI adds self-serve project leaving, dedicated Vercel API keys via TOGETHER_API_KEY, and new serverless/dedicated models.

    └──▷ HOW TO FIND IT
    Isolate Vercel project traffic and credentials by letting Together AI auto-provision a scoped API key — verify the injected variable in your Vercel project after linking.
    📍In the Together AI console, go to Integrations settings, connect your Vercel project, then confirm TOGETHER_API_KEY appears in your Vercel project's environment variables.
    Target the new high-efficiency Qwen3 serverless model in an API call to benefit from FP4 quantization at low cost.
    $ curl https://api.together.xyz/v1/chat/completions \
  -H 'Authorization: Bearer $TOGETHER_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{"model": "Qwen/Qwen3.8-2.4T-A95B", "messages": [{"role": "user", "content": "Explain buffer overflows."}]}'
    • ›Connecting a Vercel project via Integrations settings now creates a dedicated API key per linked Vercel project, automatically set as the TOGETHER_API_KEY environment variable in that Vercel project.
    • ›Project collaborators can now leave a project themselves from 'Settings > Project > Collaborators' or from the Projects list in Organization Settings — with a guard preventing the last admin from leaving before promoting another collaborator.
    • ›Adds Qwen/Qwen3.8-2.4T-A95B (FP4 quantization) to serverless, priced at $2.50 input / $6.25 output / $0.50 cached input per 1M tokens.
    • ›Adds nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-FP8 as a new model available for deployment on dedicated endpoints.
    └──▷ BREAKING ON UPGRADE
    • !nvidia/Nemotron-3-ultra-550b-a55b, pearl-ai/gemma-4-31b-it, deepseek-ai/DeepSeek-V4-Pro, and moonshotai/Kimi-K2.7-Code have been deprecated and are no longer available on serverless; use deepseek-ai/DeepSeek-V4-Pro-0813 in place of deepseek-ai/DeepSeek-V4-Pro.
  19. Tokenized dataset download in the fine-tuning console Aug 10, 2026 · issue 002

    Together AI adds tokenized dataset download, non-interactive fine-tune deletion, and project visibility controls

    └──▷ TRY IT
    Delete a fine-tuning job from CI without being prompted for confirmation.
    $ tg fine-tuning delete <job-id> --non-interactive --json
    • ›Adds tokenized dataset archive download via the fine-tuning jobs dashboard: when a job has a tokenized dataset, a 'Tokenized dataset' row with a 'Download' button appears in job details, opening a presigned archive URL; the same archive is also available via the API and CLI.
    • ›tg fine-tuning delete now honors global non-interactive mode — --non-interactive, --json, and non-TTY sessions skip the confirmation prompt, removing the need for --force in CI/scripts.
    • ›Projects now support three visibility levels — 'Open' (any org member can discover and join), 'Closed' (discoverable, join requires admin grant), and 'Private' (visible only to existing collaborators and org admins) — settable at creation or via Project Settings.
    • ›Clusters, fine-tuned models, endpoints, evaluations, files, and API keys are now fully scoped to projects, lifting the earlier early-access limitations on project isolation.
  20. Longer context for GLM-5.2 Aug 8, 2026 · issue 002

    GLM-5.2 context doubles to 512K tokens; projects gain three visibility levels and full multi-resource scoping.

    • ›Extends zai-org/GLM-5.2 serverless context length to 512,000 tokens, up from 262,144, at unchanged pricing.
    • ›Adds three project visibility levels — Open (any org member can discover and join), Closed (discoverable but join requires admin approval), and Private (visible only to existing collaborators and org admins) — configurable at creation or any time via Project Settings.
    • ›Fully scopes clusters, fine-tuned models, endpoints, evaluations, files, and API keys to projects, removing early-access limitations on project isolation.
  21. Fine-tune tokenized dataset download Aug 6, 2026 · issue 002

    Together AI adds tokenized dataset download, three-tier project visibility, and full multi-project resource scoping.

    └──▷ TRY IT
    Download the tokenized dataset for a fine-tuning job to inspect preprocessing output or archive training data.
    $ tg fine-tuning download-tokenized-dataset ft-abc123 --output-dir ./tokenized
    Retrieve a presigned download URL for a tokenized dataset programmatically, e.g. to pull it into a pipeline.
    $ curl -X GET https://api.together.ai/fine-tunes/ft-abc123/download-tokenized-dataset \
  -H 'Authorization: Bearer $TOGETHER_API_KEY'
    • ›New GET /fine-tunes/{id}/download-tokenized-dataset endpoint returns a presigned URL to download the tokenized dataset archive generated for a fine-tuning job.
    • ›New tg fine-tuning download-tokenized-dataset [FT_ID] --output-dir CLI subcommand downloads the tokenized dataset archive to a local directory.
    • ›Projects now support three visibility levels — Open (any org member can discover and join), Closed (discoverable but join requires admin approval), and Private (visible only to existing collaborators and org admins) — selectable at creation or changed anytime via Project Settings.
    • ›Clusters, fine-tuned models, endpoints, evaluations, files, and API keys are now fully scoped to projects, removing early-access limitations on project isolation.
  22. Models page visibility filter Jul 29, 2026 · issue 002

    Together AI console gains project visibility levels, cross-project model filtering, and full multi-project resource scoping.

    • ›Adds a Visibility filter on the Models page to show Internal models, Private models, or both, now pulling Internal-visibility models from every project in your organization under 'My models' — not just the selected project.
    • ›Brings Fine-tuning, Files, and Evaluations into the Projects UI in the console, allowing creation and management of these resources within a project without relying solely on project-scoped API keys.
    • ›Introduces three project visibility levels — Open (any org member can discover and join), Closed (discoverable but join requires admin approval), and Private (visible only to existing collaborators and org admins) — configurable at creation time or later via Project Settings.
    • ›Extends full project scoping to clusters, fine-tuned models, endpoints, evaluations, files, and API keys, removing the previous early-access limitations on project isolation.
  23. A/B variant percent updates in the CLI Jul 28, 2026 · issue 002

    Together AI CLI gains A/B traffic control, smart replica inference, upgrade notices, and projects get full multi-resource scoping.

    └──▷ TRY IT
    Gradually ramp a variant's traffic to 20% in a live A/B experiment without touching other variants.
    $ tg beta endpoints update <endpoint-id> --ab-percent 20
    Create a new deployment in a stopped state (zero replicas) to reserve capacity without serving traffic.
    $ tg beta endpoints deploy <model-id> --min-replicas 0
    Disable the daily CLI upgrade check in a non-interactive CI environment.
    $ export TOGETHER_DISABLE_VERSION_CHECK=1
    • ›Adds --ab-percent flag to tg beta endpoints update to change a variant's traffic percentage in an existing A/B experiment; control must stay at least 1%, and --percent on tg beta endpoints ab is capped at 1–99.
    • ›Adds smart replica-bound inference to tg beta endpoints deploy: --min-replicas alone mirrors into the max (including 0 to create a stopped deployment), and --max-replicas 0 alone lowers the min to 0.
    • ›Adds upgrade-availability detection to the Together CLI, printing a notice at most once per day and offering to run the matching install command (uv, pipx, or pip) in interactive sessions; set TOGETHER_DISABLE_VERSION_CHECK=1 to disable.
    • ›Live dedicated model inference endpoints now appear under 'My Endpoints' in the evaluations console model picker; legacy dedicated endpoints appear under 'My Legacy Endpoints'.
    • ›Projects now support three visibility levels — Open, Closed, and Private — configurable at creation or any time from Project Settings.
    +1 moreshow less
    • ›Clusters, fine-tuned models, endpoints, evaluations, files, and API keys are now fully scoped to projects, lifting all early-access isolation limitations.
    └──▷ BREAKING ON UPGRADE
    • !On tg beta endpoints update, stopping a deployment requires both --min-replicas 0 and --max-replicas 0; passing a single zero bound is now an error.
  24. New serverless models Jul 27, 2026 · issue 002

    Together AI adds Kimi-K3 serverless model with 1M context, Qwen3.6-27B fine-tuning, and project visibility controls.

    • ›Adds moonshotai/Kimi-K3 as a serverless model with 1,000,000-token context length, priced at $3.00 input / $15.00 output / $0.30 cached input per 1M tokens, with support for function calling, structured outputs, and vision inputs.
    • ›Adds Qwen/Qwen3.6-27B as a supported model for fine-tuning.
    • ›Adds three project visibility levels — Open (any org member can discover and join), Closed (discoverable but join requires admin approval), and Private (visible only to existing collaborators and org admins) — configurable at project creation or via Project Settings.
    • ›Clusters, fine-tuned models, endpoints, evaluations, files, and API keys are now fully scoped to projects, removing early-access limitations on project isolation.
my-toolchain — 0 tools
paste an install list to detect your tools

A brew list, a Brewfile, requirements.txt, a Dockerfile — or just the product names, free-form. Nothing leaves your browser.

    browse all tools →