'%3e%3crect%20x='16'%20y='51'%20width='68'%20height='27'%20rx='13.5'/%3e%3ccircle%20cx='32'%20cy='53'%20r='15'/%3e%3ccircle%20cx='53'%20cy='43'%20r='20'/%3e%3ccircle%20cx='71'%20cy='57'%20r='12'/%3e%3c/g%3e%3c/svg%3e){kind=small}
Cloud Changelog
Together AI platform changelog
272 announcements since July 2025 · 1 GA
Free. Unsubscribe any time.
RSS feed for Together AI platform
November 202610
Added to serverless with 262,144 context, FP8, function calling, structured outputs
Messages and Responses APIs add GLM-5.3 and DeepSeek-V4.1-Flash models
moonshotai/Kimi-K3.1-Pro on serverless, with 1,048,576-token context
MiniMax M3.1 Flash adapters can be deployed on dedicated model inference
Together Python SDK and CLI 3.1.0 add expiring secrets and GB200 clusters
MiniMax-M3 deprecated for new post-training resources, use MiniMax-M3.1-Flash
Anthropic Messages API and OpenAI Responses API supported on Together AI
October 202628
MiniMax-M3.1-Flash available for LoRA post-training with 262,144 context
Llama-3.3-70B-Instruct-Turbo and DeepSeek-V4-Flash-0731 deprecated on serverless
MiniMax-M3.1-Flash on serverless with 1,048,576-token context window
Billing usage line items add applied_credit object, split by credit type
Together Python SDK 3.0, TypeScript SDK 1.0 and CLI 3.0 released with breaking changes
Kimi-K3 post-training context raised to 163,840 tokens from 65,536
DeepSeek-V4.1-Flash-1017 on serverless, with a 1M-token context window
Muse-Spark-Open-816B-A42B available for LoRA post-training on B200-SXM
FP4 preparation uses model_object_id; adapter_object_id names deprecated
reflection/Beam-1.0-501B-A23B on serverless, with a 1,048,576-token context
Batch jobs now run against a single model; mixed-model files fail validation
US-only serverless inference via us-api.together.ai, at 1.25× standard rates
reflection/Beam-1.0-501B-A23B available on dedicated endpoints
NVIDIA-Nemotron-3.5-Super-VL available on dedicated endpoints
Muse-Spark-Open-816B-A42B and Nemotron-3.5-Super-VL available on serverless
Kimi-K3 price rises to $3.00 input and $15.00 output per 1M tokens
Full fine-tuning max batch size for Qwen 27B models lowered to 4
Parakeet TDT 0.6B v3 and Whisper Large v3 available on dedicated endpoints
Post-training supports custom training loops with SFT, DPO and GRPO, billed per GPU-minute
Together CLI 2.41.0 adds browser-based tg login and tg switch
GLM-5.3, GLM-5.3-Flash and DeepSeek-V4.1-Flash on provisioned throughput
Serverless dynamic rate limits that scale with recent traffic removed
Together Link in beta runs six coding agents on Together AI models
together-sandbox SDK for Python and TypeScript, available on an allowlist
Kokoro-82M and Qwen3-TTS-12Hz-1.7B-CustomVoice on dedicated endpoints
September 202619
LoRA fine-tuning supports rank up to 128, default stays at 64
Qwen 27B models support 131,072-token SFT context, with lower batch size limits
Together CLI 2.24.0 deploys fine-tuned models by model_object_name
Tev1-4B-experimental available on serverless with 32,768 context
Qwen3.7-Max and Qwen3.8-Flash prices lowered, effective September 22, 2026
Deployments scale to zero after an inactivity timeout you set
Rollouts shift endpoint traffic between deployments with canary, blue-green or rolling
Together CLI and Python SDK 2.34.0 adds rollouts, rejects oversized Parquet files
gpt-oss-20b, gemma-4-31B-it and two other models deprecated on serverless
Fine-tuning priced per model instead of parameter-count tier, many models cheaper
Three models deprecated for fine-tuning, including Nemotron-3 and Qwen3.5-35B-A3B-Base
DeepSeek-V4.1-Flash on serverless with 1M context, $0.30 input per 1M tokens
LoRA fine-tuning no longer targets expert layers of FP4, MXFP4 and NVFP4 MoE models
Together CLI 2.33.2 shows API error messages and upload progress bars
Preemptible compute in public preview for Kubernetes GPU clusters
Four serverless models deprecated, removal on September 14, 2026
August 202650
Qwen3.8-Flash on serverless with 1M context, at $0.15 input / $0.47 output per 1M tokens
Dedicated endpoint autoscaling adds active_sessions metric for TTS WebSocket sessions
Notify when available dialog adds GPU type, region, dates and estimated cost
GLM-5.3 on serverless with 1M context, at $1.40 input / $4.40 output per 1M tokens
Scheduled capacity clusters can start immediately with Start now in the console
GET /billing/usage endpoint in beta for organization usage and cost
Legacy API keys can be regenerated by default project admins and editors
Four models deprecated on serverless, including DeepSeek-V4-Pro and Kimi-K2.7-Code
tg files upload and check show a progress bar; Python SDK adds progress_callback
Project members can download kubeconfig for Slurm clusters from cluster details
GLM-5.3-Flash on serverless, with 1M context and function calling
GLM-5.2 on serverless raises context length to 1,000,000 tokens
Quick reprovision manual node repair available for Pending or Scheduling nodes
Fine-tuning training quality improves for Qwen3.5, Qwen3.6 and Nemotron 3 models
deepcogito/cogito-v2-1-671b deprecated and removed from serverless
Console endpoint forms use Deployment profiles cards, replacing separate hardware controls
Auto node repair adds Fully automatic confirmation policy, with per-fault scoping
GPU cluster size picker lists all full-node multiples of 8 GPUs beyond 1,024
FLUX.1-schnell, Qwen2.5-7B and Kimi-K2.6 deprecated on serverless
ACH bank transfers available to all customers for purchasing credits
Scheduled capacity clusters can mix capacity pool GPUs with reserved GPUs
Dedicated endpoint management API rejects unknown fields and parameters with HTTP 400
DeepSeek-V4-Pro-0813 on serverless with 1,048,576 context, $1.32 input / $3.96 output
Multi-project enabled for every organization, with resources fully scoped to projects
tg fine-tuning model-limits CLI command shows base model capability flags and bounds
Tokenized dataset upload event now includes the CLI download command
Fine-tuning CLI adds tg ft model-limits and download-tokenized-dataset commands
tg beta endpoints events lists audit and lifecycle events for dedicated endpoints
Vercel integration creates a dedicated API key per linked Vercel project
tg beta clusters delete commands no longer hang in CI, add --force flag
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-FP8 available on dedicated endpoints
Qwen3.8-2.4T-A95B now available on serverless, in FP4 quantization
Muse-Glimmer-30B and Seedance-2.5 now available on serverless
tg fine-tuning delete skips confirmation prompt in non-interactive mode
GLM-5.2 on serverless context length raised to 512,000 tokens, price unchanged
Project API keys can be given an expiration date in the console
Ternary-Bonsai-27B, p-image-ideogram and FLUX-3 added to serverless
LoRA fine-tuning on DeepSeek-V3.1 can target MoE expert layers
Download tokenized datasets from fine-tuning jobs via API and CLI
DeepSeek-V4-Flash-0731 on serverless with 1M context, $0.14 input / $0.28 output
July 202664
Models page shows Uploading badge and live progress log for remote uploads
Inkling-Small on serverless with 524,288 context, $0.50 input / $1.20 output
OIDC Kubeconfig dialog adds cache-clear command and suggests kubectl get pods
Many Qwen, Kimi, GLM, DeepSeek and Llama models deprecated for fine-tuning
Models page lists Internal models from all projects, with a Visibility filter
Fine-tuning, Files and Evaluations added to the Projects UI in the console
tg beta endpoints update adds --ab-percent to change A/B variant traffic
tg beta endpoints deploy infers a missing replica bound from --min-replicas or --max-replicas
Evaluations console model picker lists live dedicated endpoints under My Endpoints
Together CLI prints a daily upgrade notice when a newer release is available
tg beta endpoints get accepts endpoint and deployment names as well as IDs
Kimi K3 available on serverless with a 1,000,000-token context length
Dedicated endpoint deployments over GPU quota now return HTTP 429
Fine-tuning comparison view adds Metrics filtering for sampling rate and step range
Python SDK adds client.beta.realtime.transcription() for streaming speech-to-text
Project API keys can have scheduled expiration canceled in the console
Fine-tune responses include qualified model registry names for models and checkpoints
Dedicated container inference supports HTTP server mode for OpenAI-compatible endpoints
tg fine-tuning list-checkpoints now shows registry artifact IDs
tg fine-tuning preview shows how a base model tokenizes training file rows
Dedicated endpoint metrics add KV-cache utilization and cache hit rate gauges
Node repair adds Cycle action that power-cycles bare-metal hosts through the BMC
tg beta clusters remediations approve adds --mode flag to override repair action
jig deploy rejects configs with a name in both environment variables and secrets
Pairwise InfiniBand Write Bandwidth health check added for GPU clusters
tg beta clusters create and update add flags for Headlamp and Slurm Web add-ons
Fine-tuning on MoE models can target expert feed-forward layers via lora_trainable_modules
Reserved GPU count on a reserved cluster can be increased via API and CLI
Fine-tuning dashboard adds a Metrics tab with live training curves and run comparison
DeepSeek-V4-Flash available for deployment on dedicated endpoints
Model and adapter revisions return validationStatus, lastValidatedAt and validationErrors
Privacy toggles, including training opt-in, move to Organization Settings
Endpoint get and list responses limited to the 10 newest deployment summaries
Cluster responses add num_reserved_gpus and num_capacity_pool_gpus fields
Together Python SDK 2.24 drops API key requirement for OIDC cluster SSH
Dedicated model inference adds one-command deploy, A/B tests and autoscaling
thinkingmachines/Inkling available on serverless, with 524,288 context length
Queue API adds POST /queue/clear to cancel all pending jobs for a model
Fine-tune jobs and checkpoints return Together model registry IDs
Evaluations support vision models with an image_data_urls dataset column
Remediation responses include linked_alerts for passive health check alerts
TogetherLink open-source launcher connects Claude Code, Codex and more to Together AI
Model limits endpoint returns supports_full_training for fine-tuning checks
Slurm GPU clusters with OIDC support browser-based SSH via Together CLI
Passive health checks document monitored failure signals and repair actions
google/gemma-4-12B-it available on serverless with 262,144-token context
Qwen3-235B-A22B-Instruct-2507-tput, Llama-3-8B-Instruct-Lite and GLM-5.1 deprecated on serverless
Storage performance health check added for GPU clusters, using fio
TorchTitan training health check in preview for B200 GPU clusters
LiquidAI/LFM2.5-8B-A1B available on serverless with 32,768 context length
LiquidAI/LFM2-24B-A2B deprecated on serverless, replaced by LFM2.5-8B-A1B
Provisioned throughput available for MiniMax-M3 and GLM-5.2, one-month minimum term
HappyHorse-1.0-T2V text-to-video model available on serverless
Usage object token details vary by model; read both shapes defensively
Reasoning field now symmetric for input and output on reasoning models
Project admins can change a project's slug from Project Settings
Cost analytics adds a Measure control to switch between Cost ($) and Units
Fine-tuning now supports the vision-language model google/gemma-4-31B-it-VLM
BYOM fine-tuning supports Hugging Face models built with Transformers v5 or earlier
Serverless adds Gemini 3.1 Flash-Lite Image, Qwen3.6-35B-A3B-Lora and HappyHorse 1.1 video models
June 202624
Qwen/Qwen3.5-397B-A17B deprecated and removed from serverless
Seedance 2.0 adds 4k resolution tier (up to 3840x2160) at $0.836 per second
GPU clusters add passive health checks and automatic node repair
POST /fine-tunes/estimate-price returns estimated fine-tuning job cost before launch
Fine-tuning jobs support early stopping based on validation loss
GET /whoami shows the API key, project and organization behind a request
Binary audio uploads for transcription and translation capped at 80 MB
Multiple LoRA adapters can share one dedicated endpoint, in preview
GLM-5.2 available on serverless with 262K context at $1.40 input per 1M tokens
Organization and project roles renamed to Admin, Developer and Editor labels
Qwen3.5-397B-A17B deprecated on serverless after June 29, 2026
Kimi-K2.7-Code available on serverless with 262,144 context, $0.95 input per 1M
MiniMax-M3 on serverless with 524,288-token context, $0.30 input per 1M tokens
Voxtral-Mini-3B-2507 deprecated on serverless, still available as dedicated endpoint
Cached input pricing for GLM-5.1 and Qwen3.5-397B-A17B, DeepSeek-V4-Pro price cut
Fine-tuning file uploads get server-side schema validation with processing_status
Qwen3-Coder-480B-A35B-Instruct-FP8 deprecated on serverless; MiniMax-M2.7 suggested
GET /fine-tunes/{id}/metrics endpoint returns fine-tuning training metrics
Billing docs updated for multiple payment methods and prepaid-only access
GPU clusters support Slurm startup scripts for node startup and per-job hooks
Compare evaluator adds disable_position_bias_correction, halving judge cost and latency
May 202618
Price increases for Qwen3.5-9B, Llama 3 8B Lite and Llama 3.3 70B Turbo
FLUX.1-krea-dev deprecated and no longer available on serverless
Gemma, Qwen2.5 and Llama 3.2 among 17 models added to dedicated endpoints
Quickstart added for Seedance 2.0 video generation, covering text-to-video and editing
Seedream 5.0 Lite, HappyHorse 1.0 and Veo 3.1 models now on serverless
Qwen3.7-Max available on serverless at $2.50 input / $7.50 output per 1M tokens
GPU clusters support external OIDC authentication for Kubernetes clusters
pearl-ai/gemma-4-31b-it available on serverless with 32,000 context length
DeepSeek-R1, DeepSeek-V3.1 and Qwen3-Coder-Next-FP8 deprecated on serverless
google/gemma-4-31b-it price rises to $0.39 input, $0.97 output from May 21, 2026
HappyHorse 1.0 t2v and Seedance 2.0 video models available on serverless
Audio transcription and translation endpoints accept .ogg, .opus and .aac files
Together CLI uses tg as the canonical command name, with a refreshed command tree
task field no longer required in audio verbose JSON transcription and translation responses
Slurm GPU clusters move to Slurm-on-Kubernetes stack with self-healing workers
April 202613
TTS requests add pronunciation_dict parameter for custom pronunciation rules
Deployments can autoscale on any custom Prometheus metric from /metrics
Fine-tuning adds Qwen3.6-35B-A3B, Gemma 4 31B and Gemma 4 26B-A4B
Serverless adds Cogito v2.1 671B, Vidu Q3 and Q3 Turbo, and Wan 2.7 video models
DeepSeek-V4-Pro added to serverless with 512,000-token context, $2.10 input per 1M
No-packing fine-tuning jobs now priced on len(dataset) max_seq_length tokens
Build, Scale and Enterprise tiers retired; billing fully prepaid with dynamic rate limits
google/gemma-3n-E4B-it pricing rises to $0.06 input, $0.12 output per 1M tokens
Qwen3-VL-8B-Instruct, Qwen3-235B-A22B-Thinking-2507 and Mixtral-8x7B deprecated
GLM-4.5-Air-FP8 and GLM-4.7 deprecated, along with Qwen3-Next-80B-A3B-Instruct
March 20264
February 20269
FLUX.1-dev, Qwen3-VL-32B-Instruct and 11 other models deprecated
Dedicated Container Inference (DCI), formerly BYOC, launches for custom models
Ten models deprecated, including Llama-4-Scout-17B-16E and Qwen2.5-72B-Instruct-Turbo
Python SDK v2.0 released, type-safe and OpenAPI-generated, with beta instant clusters APIs