Last updated · 97 releases tracked since December 2022
A dated log of everything Together AI has shipped, collected from official Together AI sources and updated as new releases land.
September 2026 releases included H100 dedicated endpoint pricing reduction, new serverless models with 1,000,000 context length, preemptible GPU compute preview, and rollouts feature for traffic management.
Summary of September 2026, generated from the releases listed below.
This month
7
-26 vs last month
Last 12 months
72
Busiest: September 2026
Most shipped
Infrastructure
28 releases
Newest first. Each entry links to the official page it came from.
Post-training lets you write your own training loop and run it against Together AI GPUs. Your code builds each batch and chooses the loss function and learning rate, while the GPUs you provision run the forward and backward passes, optimizer steps, and sampling. It bills per GPU…
The following models are now available on provisioned throughput: zai-org/GLM-5.3, zai-org/GLM-5.3-Flash, deepseek-ai/DeepSeek-V4.1-Flash.
Serverless no longer applies dynamic rate limits that scale with recent traffic. Requests may still be limited during high demand periods.
Beta release of Together Link, a tool that runs six coding agents on models hosted by Together AI, including Claude Code, Codex, OpenCode, and Pi Code. It offers features like auto router, cost tracking, and image generation.
Together Link runs six coding agents on models hosted by Together AI: Claude Code, Codex, OpenCode, and Pi Code in the terminal, plus Claude Desktop (including Cowork) and ChatGPT Desktop. Install it with one command, launch your agent through it, and your normal agent configura…
The following models are now available for deployment on dedicated endpoints: hexgrad/Kokoro-82M (text-to-speech), Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice (text-to-speech).
New Code sandbox SDK and CLI with Together-native authentication, Python support, and Docker-defined environments. Replaces legacy SDK with improved features and workflows.
Extended fine-tuning context for Qwen 27B models to 131,072 tokens for SFT and 65,536 for DPO, with updated batch size limits.
Increased LoRA rank limit for fine-tuning to 128 for most models, with updated CLI/SDK versions required for optimal use.
Qwen/Qwen3.8-27B, Qwen/Qwen3.6-27B, and Qwen/Qwen3.5-27B now support a 131,072-token context for SFT (up from 32,768) and 65,536 for DPO (up from 16,384), for both LoRA and full fine-tuning.
You can now train low-rank adapters (LoRA) with a rank of up to 128 for the majority of models, up from 64. The default rank for these models stays at 64, so set lora_r to use a higher one.
Batch API files now retained for 7 days after processing, with access restrictions after expiration.
Deploy fine-tuned models using registry names instead of model_object_id, simplifying deployment workflows.
Deploy a fine-tuned model by its registry name using the Together CLI version 2.24.0 or later. The CLI resolves the name to the same model, so you can deploy straight from the name shown in the fine-tuning jobs dashboard.
New serverless model with 32,768 context length and updated pricing details.
The following models are now available on serverless: together/Tev1-4B-experimental: 32,768 context length. Pricing: $0.042 input / free output (per 1M tokens).
Pricing reductions for Qwen models effective September 22, 2026, with updated cost structure for input/output tokens.
Pricing update for Qwen models effective September 22, 2026 with reduced input/output rates.
The following models have lower pricing, effective September 22, 2026. All usage from that date forward is billed at the new rates (per 1M tokens): Qwen/Qwen3.7-Max, Qwen/Qwen3.8-Flash.
Automatic idle shutdown feature for dedicated deployments, allowing self-scaling based on inactivity timeouts.
Together CLI v2.34.0 adds HIPAA placement support and rollouts functionality for dedicated deployments.
Version 2.34.0 of the Together CLI and Python SDK carries the rollouts release above (the SDK surface is client.beta.endpoints.rollouts) and also adds HIPAA placement support.
Rollouts shift live traffic from one deployment to another under the same endpoint, without changing the endpoint URL. Pick a canary, blue-green, or rolling strategy to determine how traffic moves.
The following models have been deprecated and are no longer available on serverless: openai/gpt-oss-27B, google/gemma-4-31B-it, thinkingmachines/Inkling-Small, intfloat/multilingual-e5-large-instruct.
HIPAA-compliant placement option for deployments, restricting them to HIPAA-attested clusters for compliance requirements.
Rollouts feature for shifting traffic between deployments without changing endpoint URLs, supporting canary, blue-green, and rolling strategies.
The following models are now available on serverless: MiniMaxAI/MiniMax-H3: Pricing: $0.1391/sec at 2k.
New serverless model with 2k/sec pricing and 1,000,000 context length for inference tasks.
The following models are now available on serverless: deepseek-ai/DeepSeek-V4.1-Flash: 1,000,000 context length, FP8 quantization, function calling, and structured outputs. Pricing: $0.30 input / $1.20 output / $0.006 cached input (per 1M tokens).
New serverless model with 1,000,000 context length, FP8 quantization, and function calling capabilities.
Together CLI v2.33.2 improves error reporting and upload feedback with progress bars.
Version 2.33.2 of the Together CLI improves error reporting and upload feedback: Endpoint, fine-tuning, and model commands now print the API’s error message when a request fails, rather than a generic failure notice.
Preemptible compute preview for GPU clusters with discounted rates, sub-hourly billing, and 5-minute drain windows for node reclamation.
Preemptible compute is now in public preview for Kubernetes GPU clusters. Alongside standard nodes, you can set a preemptible GPU target, and Together provisions toward it as spare capacity becomes available, at a flat discounted rate relative to on-demand.
New serverless model available for fine-tuning with 1,000,000 context length and FP4 quantization.
You can now fine-tune the following models: zai-org/GLM-5.3.
Deprecation of openai/gpt-oss-20b and other models from serverless, with recommended replacements and migration guidance.
H100 80GB dedicated endpoint hardware is now $3.99 per hour, down from $5.49. See Dedicated endpoint pricing.
The following models are deprecated and will be removed from serverless on September 14, 2026: openai/gpt-oss-20b, google/gemma-4-31B-it, thinkingmachines/Inkling-Small, intfloat/multilingual-e5-large-instruct.
Pricing reduction for H100 80GB dedicated endpoints from $5.49 to $3.99 per hour.
New serverless model with 1,000,000 context length and updated pricing structure.
The following models are now available on serverless: Qwen/Qwen3.8-Flash: 1,000,000 context length. Pricing: $0.15 input / $0.47 output (per 1M tokens).
TTS autoscaling metric support for active_sessions, enabling better scaling based on concurrent WebSocket sessions.
New serverless model with 1,000,000 context length, FP4 quantization, and structured output capabilities.
The following models are now available on serverless: zai-org/GLM-5.3: 1,000,000 context length, FP4 quantization, function calling and structured outputs. Pricing: $1.40 input / $4.40 output / $0.26 cached input (per 1M tokens).
Project members can now download the cluster kubeconfig for Slurm clusters from the cluster details page, matching Kubernetes cluster behavior.
Slurm cluster kubeconfig access for project members, aligning with Kubernetes cluster behavior for cluster management.
Deprecation of multiple models from serverless, with recommended replacements and migration options provided.
The following models have been deprecated and are no longer available on serverless: nvidia/Nemotron-3-ultra-550b-a55b, pearl-ai/gemma-4-31b-it, deepseek-ai/DeepSeek-V4-Pro, moonshotai/Kimi-K2.7-Code.
Legacy API key regeneration capability for project admins/editors, with deprecated badges for legacy keys.
tg files upload and tg files check now show a progress bar in interactive terminals. The Python SDK accepts an optional progress_callback on client.files.upload() that receives upload progress events.
File upload progress tracking in CLI and Python SDK with progress bars and callback support for monitoring uploads.
Beta Billing Usage API for retrieving organization-wide usage and cost data, filterable by month.
New serverless model available for fine-tuning with updated technical specifications and pricing.
The Together CLI now includes a tg batches command group for batch inference: tg batches submit uploads a local JSONL file (or takes an existing file ID) and creates a job against chat.completions, audio.transcriptions, or audio.translations.
New batch jobs CLI commands for managing batch inference workflows with submit, list, retrieve, and download capabilities.
Kimi K3, an open-weight model, is available on Together AI with lower costs and improved pass@k metrics compared to GPT-5.6 Sol.
Aurora is an open-source, RL-based framework that learns from live inference traces to continuously update the speculator without interrupting serving, achieving a 1.25x speedup over static speculators.
Together AI introduced cached input token pricing for MiniMaxAI/MiniMax-M2.5, reducing cached token costs on serverless inference.
Together AI added Qwen/Qwen3.5-9B to serverless inference, expanding the model set available to builders on the platform.
Together AI deprecated several older models in March, including Moonshot Kimi models and legacy rerank and embedding endpoints.
Together AI added Qwen3.5-397B-A17B to serverless inference, widening access to a larger reasoning-focused model.
MiniMax-M2.5 joined Together AI's serverless lineup, adding another February model release for chat and agentic workloads.
Together AI brought GLM-5 into serverless inference, expanding its February model catalog with a fresh general-purpose release.
Together AI launched Dedicated Container Inference for custom models, letting teams deploy containerized workloads on managed infrastructure.
Together AI's Python SDK v2.0 reached general availability with a type-safe, OpenAPI-driven client and a migration guide from the legacy v1 SDK.
Together AI added Qwen/Qwen3-Coder-Next-FP8 to serverless inference, giving developers another newer coding model on the platform.
Together AI added Kimi-K2.5 to serverless inference, broadening access to Moonshot AI's multimodal reasoning model.
Together AI made prompt caching the default on newly created Dedicated Endpoints and deprecated the opt-out flag.
Together AI added Cartesia Sonic-3, a low-latency text-to-speech model served through its API for high-fidelity voice generation.
Together AI added GLM-4.7 to serverless inference, giving builders direct access to Z.ai's latest reasoning and agentic coding model.
Qwen3-VL-32B-Instruct is the most powerful vision-language model in the Qwen series with 33B parameters, 256K native context, and advanced visual agent capabilities for PC/mobile GUI operations.
Together AI introduces a fine-tuning platform for OpenAI's gpt-oss-120B and gpt-oss-20B models, enabling domain-specific customization with LoRA training and enterprise deployment capabilities.
Qwen Image delivers reliable image generation capabilities with excellent cost-efficiency for diverse applications.
Together AI achieved SOC 2 Type 2 compliance, demonstrating rigorous security controls across infrastructure, including access management, data encryption, and incident response protocols.
Together Code Sandbox provides a fast, reliable infrastructure for building and scaling AI coding tools with features like instant VM startup, scalable sandboxes, and advanced developer tools.
Together AI now enables deployment of NVIDIA NIM models with enterprise-grade infrastructure, including 40+ models for production workloads and integration with NVIDIA API Catalog.
Announces the availability of DeepSeek-R1 and distilled models with secure deployment options on Together AI's platform
Qwen 14B distilled with reasoning capabilities from Deepseek R1. Outperforms GPT-4o in math & matches o1-mini on coding.
DeepSeek R1 Distilled Llama 70B is a 70B parameter language model with advanced reasoning capabilities, surpassing GPT-4o on MATH-500 (94.5%) and matching o1-mini on coding benchmarks. Designed for researchers and developers requiring high-performance language models.
Together AI acquires CodeSandbox to launch the Together Code Interpreter, enabling code execution within generative AI workflows.
Introduces LoLCATs, a method for creating subquadratic LLMs with linear-time and constant-memory generation, enabling efficient model linearization.
NVIDIA NIM for GPU accelerated Llama 3.1 Nemotron 70B Instruct inference through OpenAI compatible APIs. Fine-tuned for alignment and helpfulness, providing accurate and informative responses. Suitable for developers and researchers requiring advanced language understanding capa…
Introducing the Together Enterprise Platform for secure GenAI deployment with 2x faster inference, continuous model optimization, and enterprise-grade GPU orchestration.
Announced serverless Together Rerank API and exclusive access to Salesforce LlamaRank model for enterprise search enhancement.
Introduces SpecExec, a new speculative decoding method for accelerating LLM inference on consumer devices through parallel processing and optimized tree structures.
Introduces Mixture of Agents (MoA), a layered approach using multiple open-source LLMs to enhance performance, achieving 65.1% on AlpacaEval 2.0.
Introduces Mixture of Agents (MoA), an approach leveraging multiple open-source LLMs to achieve state-of-the-art performance with 65.1% on AlpacaEval 2.0.
Together AI partners with Meta to release Meta Llama 3 for inference and fine-tuning, including models like Llama-3-8B-Instruct, Llama-3-70B-Instruct, Llama-3-8B, and Llama-3-70B, with performance up to 350 tokens per second via Together Inference Engine 2.0.
Launch of the Together Embeddings endpoint for building RAG applications with Langchain
Mistral Instruct generates human-like text based on user input, excelling at conversational tasks with its instruction fine-tuning.
Announced partnership with Y Combinator for dedicated GPU clusters, offering faster GPU access for YC startups.
Together AI and Y Combinator launch dedicated GPU clusters for YC community with faster GPU access
Together AI launches Together API and Together Compute, cloud services offering a full stack solution for developers to train, fine-tune, and run open-source AI models.
The RedPajama base dataset is a 1.2 trillion token fully-open dataset created by following the recipe described in the LLaMA paper. It includes data slices like CommonCrawl, C4, GitHub, arXiv, Wikipedia, and StackExchange, processed with quality filters to match the LLaMA datase…
Released a 7B parameter, 8-bit quantized model for consumer GPUs, improved 20B model with user feedback fine-tuning, new open-source dataset 'together-user-feedback', multi-GPU support, and training scripts.
Introduces H3 and FlashConv for improving state space models in language modeling, achieving results comparable to Transformers.
Releases per month over the last 12 months. Busiest month: September 2026, 33 releases.
Releases in the last 12 months by category and product.
The most recent Together AI update we track is “Post-training”, released on October 13, 2026.
Together AI shipped 7 updates in October 2026, including “Post-training”, “New provisioned throughput models”, “Dynamic rate limits for serverless retired” and 4 mores.
We tracked 72 updates from Together AI in the last 12 months, about 6 a month. The busiest month was September 2026, with 33 releases.
Together AI publishes release notes on Together AI Changelog and Together AI Docs Changelog, linked at the top of this page. Every entry here links back to the official page it came from.
Newsletter
This viewscreen turns the calendar into a recurring digest: the sharpest launches, model drops, and product moves delivered in one clear read.
The standout launches, model drops, and platform moves worth tracking.
One tight read instead of bouncing between company blogs and timelines.
Clear summaries you can scan fast and forward to the team when needed.
Monthly dispatch
Get a clean monthly dispatch covering the launches that mattered, not every minor product breadcrumb.
One email each month, tuned for signal over noise.
