Deprecations

The following models have been deprecated and are no longer available for fine-tuning:

• nvidia/NVIDIA-Nemotron-Nano-9B-v2.
• Qwen/Qwen3-Next-80B-A3B-Instruct.
• Qwen/Qwen3-Next-80B-A3B-Thinking.
• Qwen/Qwen3-0.6B.
• Qwen/Qwen3-0.6B-Base.
• Qwen/Qwen3-1.7B.
• Qwen/Qwen3-1.7B-Base.
• Qwen/Qwen3-4B.
• Qwen/Qwen3-4B-Base.
• Qwen/Qwen3-8B.
• Qwen/Qwen3-8B-Base.
• Qwen/Qwen3-14B.
• Qwen/Qwen3-14B-Base.
• Qwen/Qwen3-32B.
• Qwen/Qwen3-30B-A3B-Base.
• Qwen/Qwen3-30B-A3B.
• Qwen/Qwen3-30B-A3B-Instruct-2507.
• Qwen/Qwen3-235B-A22B.
• Qwen/Qwen3-235B-A22B-Instruct-2507.
• Qwen/Qwen3-Coder-30B-A3B-Instruct.
• Qwen/Qwen3-Coder-480B-A35B-Instruct.
• Qwen/Qwen3-VL-8B-Instruct.
• Qwen/Qwen3-VL-32B-Instruct.
• Qwen/Qwen3-VL-30B-A3B-Instruct.
• Qwen/Qwen3-VL-235B-A22B-Instruct.
• Qwen/Qwen2.5-72B-Instruct.
• Qwen/Qwen2.5-72B.
• Qwen/Qwen2.5-32B-Instruct.
• Qwen/Qwen2.5-32B.
• Qwen/Qwen2.5-14B-Instruct.
• Qwen/Qwen2.5-14B.
• Qwen/Qwen2.5-7B-Instruct.
• Qwen/Qwen2.5-7B.
• Qwen/Qwen2.5-3B-Instruct.
• Qwen/Qwen2.5-3B.
• Qwen/Qwen2.5-1.5B-Instruct.
• Qwen/Qwen2.5-1.5B.
• Qwen/Qwen2-72B-Instruct.
• Qwen/Qwen2-72B.
• Qwen/Qwen2-7B-Instruct.
• Qwen/Qwen2-7B.
• Qwen/Qwen2-1.5B-Instruct.
• Qwen/Qwen2-1.5B.
• moonshotai/Kimi-K2.5.
• moonshotai/Kimi-K2-Thinking.
• moonshotai/Kimi-K2-Instruct-0905.
• moonshotai/Kimi-K2-Instruct.
• moonshotai/Kimi-K2-Base.
• zai-org/GLM-5.
• zai-org/GLM-4.7.
• zai-org/GLM-4.6.
• deepseek-ai/DeepSeek-R1-0528.
• deepseek-ai/DeepSeek-R1.
• deepseek-ai/DeepSeek-V3-0324.
• deepseek-ai/DeepSeek-V3.
• deepseek-ai/DeepSeek-V3.1-Base.
• deepseek-ai/DeepSeek-V3-Base.
• deepseek-ai/DeepSeek-R1-Distill-Llama-70B.
• deepseek-ai/DeepSeek-R1-Distill-Llama-70B-32k.
• deepseek-ai/DeepSeek-R1-Distill-Llama-70B-131k.
• deepseek-ai/DeepSeek-R1-Distill-Qwen-14B.
• deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B.
• meta-llama/Llama-4-Scout-17B-16E.
• meta-llama/Llama-4-Maverick-17B-128E.
• meta-llama/Llama-3.3-70B-32k-Instruct-Reference.
• meta-llama/Llama-3.3-70B-131k-Instruct-Reference.
• meta-llama/Llama-3.2-3B-Instruct.
• meta-llama/Llama-3.2-3B.
• meta-llama/Llama-3.2-1B-Instruct.
• meta-llama/Llama-3.2-1B.
• meta-llama/Meta-Llama-3.1-8B-131k-Instruct-Reference.
• meta-llama/Meta-Llama-3.1-8B-Reference.
• meta-llama/Meta-Llama-3.1-8B-131k-Reference.
• meta-llama/Meta-Llama-3.1-70B-Instruct-Reference.
• meta-llama/Meta-Llama-3.1-70B-32k-Instruct-Reference.
• meta-llama/Meta-Llama-3.1-70B-131k-Instruct-Reference.
• meta-llama/Meta-Llama-3.1-70B-Reference.
• meta-llama/Meta-Llama-3.1-70B-32k-Reference.
• meta-llama/Meta-Llama-3.1-70B-131k-Reference.
• meta-llama/Meta-Llama-3-8B-Instruct.
• google/gemma-3-270m.
• google/gemma-3-270m-it.
• google/gemma-3-1b-it.
• google/gemma-3-1b-pt.
• google/gemma-3-4b-it.
• google/gemma-3-4b-it-VLM.
• google/gemma-3-4b-pt.
• google/gemma-3-12b-it.
• google/gemma-3-12b-it-VLM.
• google/gemma-3-12b-pt.
• google/gemma-3-27b-it.
• google/gemma-3-27b-it-VLM.
• google/gemma-3-27b-pt.
• mistralai/Mixtral-8x7B-v0.1.
• mistralai/Mistral-7B-Instruct-v0.2.
• mistralai/Mistral-7B-v0.1.
• togethercomputer/llama-2-7b-chat.

See Deprecations for migration options.

Improvements

Deprecations

A/B variant percent updates in the CLI

tg beta endpoints update now accepts --ab-percent to change a variant’s traffic percentage in an existing A/B experiment. The flag takes percentage from or returns it to the control only; other variants stay unchanged. The control must remain at least 1%, and --percent on tg beta endpoints ab is limited to 1–99.
See Ramp the variant and the endpoints CLI reference.

Deploy replica bound inference

tg beta endpoints deploy now infers a missing replica bound: --min-replicas alone mirrors into the max (including 0 to create a deployment stopped), and --max-replicas 0 alone lowers the min to 0. On tg beta endpoints update, stopping a deployment still requires both --min-replicas 0 and --max-replicas 0. Passing a single zero bound is an error.
See the endpoints CLI reference.

CLI upgrade notices

The Together CLI now detects when a newer release is available and prints an upgrade notice at most once per day. Interactive sessions offer to run the upgrade in place, using the command that matches your install (uv, pipx, or pip). Set TOGETHER_DISABLE_VERSION_CHECK=1 to turn the check off.
See Get started.

Model deprecations

The following models have been deprecated and are no longer available on serverless:
MiniMaxAI/MiniMax-M2.7.
See Deprecations for migration options.

New releases

Improvements

Python SDK realtime transcription

The Together Python SDK now includes client.beta.realtime.transcription() for streaming speech-to-text over WebSocket. Install with pip install "together[realtime]". The session reconnects with audio replay on transient drops, exposes normalized events such as TranscriptDelta and TranscriptCompleted, and supports application-level failover across endpoints with RealtimeConnectionError (code="no_healthy_workers") and session.pending_audio().

See Streaming transcription.

Fine-tuning comparison metrics filtering

The fine-tuning comparison view now includes the same Metrics filtering control as the single-job Metrics tab. Adjust Sampling rate and Step range, then click Apply to re-fetch metrics for every selected job with matching filters.

See View metrics in the dashboard.

Improvements

Dedicated containers OpenAI-compatible endpoints

Dedicated container inference now supports HTTP server mode for synchronous, OpenAI-compatible endpoints. Run your worker without the --queue flag, wire an OpenAI route with Sprocket, and call it with the OpenAI SDK or plain HTTP, with no Together-specific request shapes.

See Serve an OpenAI-compatible endpoint.

Fine-tuning output object names

Fine-tune retrieve and list-checkpoints responses now include qualified Together model registry names alongside object IDs. On GET /fine-tunes/{id}, use model_object_name and adapter_object_name (LoRA jobs) for the final artifacts in / form. On GET /fine-tunes/{id}/checkpoints, each entry adds object_name with the same naming pattern (including - or -adapter suffixes). Names are resolved on retrieve only, not on list jobs. If the project slug cannot be resolved, the name field falls back to the object ID.

On a completed job in the fine-tuning jobs dashboard, Output model shows the same qualified model_object_name and links to the registry model page.

See Model registry object IDs.

Fine-tuning checkpoint CLI output

tg fine-tuning list-checkpoints now displays registry artifact IDs in the table output. The Registry Artifact column shows object_id@object_revision_id, and a copyable Registry artifacts block prints below the table.

See List checkpoints.

Fine-tuning preview CLI command

tg fine-tuning preview samples rows from an uploaded JSONL training file and shows how a base model tokenizes them before you start a job. The table output highlights masked tokens, trained spans, and truncation; pass --json for the full response.

See Preview.

Worker engine cache metrics

Dedicated endpoint Prometheus metrics now expose worker_engine_kv_cache_utilization and worker_engine_cache_hit_rate gauges. Use them to track KV-cache pressure and prefix-reuse efficiency on deployment workers.

See Monitor endpoints and deployments.

Improvements

GPU cluster add-on CLI flags

tg beta clusters create and tg beta clusters update now expose flags for Headlamp and Slurm Web cluster add-ons.

Create: Pass --headlamp-addon or --slurm-web-addon to enable an add-on at cluster creation.

Update: Pass --headlamp / --no-headlamp or --slurm-web / --no-slurm-web to toggle add-ons on an existing cluster.

See the clusters CLI reference.

MoE expert-LoRA target modules

Fine-tuning jobs on supported mixture-of-experts models can now target expert feed-forward layers instead of attention projections. Set lora_trainable_modules to w_up,w_gate,w_down to train a compact shared-factor adapter across experts, useful when adapting domain knowledge rather than attention patterns.

See Target MoE expert layers.

Jig environment variable collision validation

jig deploy now rejects configs where the same name appears in both [tool.jig.deploy.environment_variables] and secrets. The error lists each colliding name so you can remove the duplicate or unset the secret before redeploying.

See Secrets.

New models available for fine-tuning

You can now fine-tune the following models:

• zai-org/GLM-5.1.
• zai-org/GLM-5.

See Supported models for the full list.

New releases

New models

Improvements

Fine-tuning metrics dashboard

Fine-tuning jobs now have a Metrics tab in the web dashboard that charts training progress. Metrics stream in live while a job runs and stay available after it completes.

What’s new:

• Live training curves: Watch training loss, learning rate, and gradient norm update as the job trains, without leaving the dashboard.
• Preference-tuning metrics: DPO and other preference jobs add reward accuracy, reward margin, chosen and rejected rewards and log-probabilities, and approximate KL.
• Interactive charts: Zoom and pan across all charts on a shared step axis, switch the x-axis between step and time, toggle linear or logarithmic scales, and sync a hover crosshair across every metric.
• Compare runs: Overlay metrics from multiple jobs to compare fine-tuning runs side by side.

New dedicated endpoint models

The following models are now available for deployment on dedicated endpoints:

• deepseek-ai/DeepSeek-V4-Flash.

New releases

New models

New serverless models

The following models are now available on serverless:
thinkingmachines/Inkling: 552.8B parameters, 524,288 context length, NVFP4 quantization. Pricing: $1.00 input / $4.05 output / $0.17 cached input (per 1M tokens).

Dedicated model inference

Dedicated model inference (DMI) serves a model on reserved GPUs, giving you higher throughput, lower latency, and predictable performance with no hard rate limits. It uses the same inference APIs as serverless, so you can prototype on serverless and deploy on dedicated hardware without changing your application code.

What’s new:

• Deploy in one command: tg beta endpoints deploy creates an endpoint, attaches a deployment, and routes all traffic to it.
• New resource model: Compose models, configs, endpoints, deployments, and traffic splits to control how your model is served. See Concepts.
• A/B tests and shadow experiments: Split live traffic across variants, or mirror traffic to a new deployment without serving its responses.
• Autoscaling and serving controls: Scale on request or GPU metrics, scale to zero on idle, and tune serving behavior per deployment. See Configure autoscaling.

Note

If you’re already using dedicated endpoints, you can migrate to the v2 API by following the migration guide.

New serverless models

The following models are now available on serverless:
thinkingmachines/Inkling: 552.8B parameters, 524,288 context length, NVFP4 quantization. Pricing: $1.00 input / $4.05 output / $0.17 cached input (per 1M tokens).

Image inputs for evaluations

You can now evaluate vision-capable models on image datasets. Add an image_data_urls column to your evaluation dataset — a base64 image data URL, or a list of them — and the images are attached to the model and judge requests alongside the text prompt.

See Prepare your dataset for details.

Improvements

Privacy and training opt-in settings

Training opt-in and other privacy toggles now live under Privacy in Organization Settings. Organization admins control data-sharing settings for all traffic sent under the organization’s API keys.

See Privacy and security.

Dedicated endpoint deployment listing

Endpoint get and list responses now include at most the 10 newest deployment summaries per endpoint. Use the deployments list API to retrieve every deployment on an endpoint.

See Manage endpoints and deployments.

New releases

New models

Deprecations

Improvements

Dedicated model inference

Dedicated model inference (DMI) serves a model on reserved GPUs, giving you higher throughput, lower latency, and predictable performance with no hard rate limits. It uses the same inference APIs as serverless, so you can prototype on serverless and deploy on dedicated hardware without changing your application code.

What’s new:

• Deploy in one command: tg beta endpoints deploy creates an endpoint, attaches a deployment, and routes all traffic to it.
• New resource model: Compose models, configs, endpoints, deployments, and traffic splits to control how your model is served. See Concepts.
• A/B tests and shadow experiments: Split live traffic across variants, or mirror traffic to a new deployment without serving its responses.
• Autoscaling and serving controls: Scale on request or GPU metrics, scale to zero on idle, and tune serving behavior per deployment. See Configure autoscaling.

Note

If you’re already using dedicated endpoints, you can migrate to the v2 API by following the migration guide.

New serverless models

The following models are now available on serverless:
thinkingmachines/Inkling: 552.8B parameters, 524,288 context length, NVFP4 quantization. Pricing: $1.00 input / $4.05 output / $0.17 cached input (per 1M tokens).

Image inputs for evaluations

You can now evaluate vision-capable models on image datasets. Add an image_data_urls column to your evaluation dataset — a base64 image data URL, or a list of them — and the images are attached to the model and judge requests alongside the text prompt.

See Prepare your dataset for details.

Fine-tuning artifact registry IDs

Fine-tune job and checkpoint responses now include Together model registry IDs for completed artifacts. Use model_object_id and model_object_revision_id on completed jobs, adapter_object_id and adapter_object_revision_id on LoRA jobs, and object_id and object_revision_id on list checkpoints entries to reference weights in the model registry. The Together Python SDK (2.24+) and Together TypeScript SDK expose these fields on retrieve and list-checkpoints responses.

Together Python SDK 2.24

Together Python SDK 2.24 is available. OIDC cluster SSH (tg beta clusters ssh) no longer requires a Together API key, and the bastion-to-target SSH hop skips host key prompts for ephemeral cluster nodes. See SSH into a cluster.

New releases

Improvements

TogetherLink

TogetherLink is an open-source launcher that connects Claude Code, Codex, ChatGPT Desktop, Pi Code, and OpenCode to models hosted by Together AI. Install with one command and run your existing tools without hand-editing provider settings.

See Configure Claude Code, Codex, and ChatGPT with Together AI models.

Slurm cluster OIDC SSH

Slurm GPU clusters with OIDC enabled now support browser-based SSH through the Together CLI. Choose OIDC on the cluster details page to sign in without uploading an SSH key. Key-based SSH remains available.

See Cluster management.

Fine-tuning model limits API

The model limits endpoint now returns supports_full_training so you can check whether a model supports full fine-tuning before submitting a job. When the field is false, the model is LoRA-only.

See LoRA vs. full fine-tuning.