Baseten Updates & Release Notes

joey-1 is now available through Baseten Model APIs, and its weights are open source on Hugging Face. It's a 2B-parameter decision model post-trained from Qwen3.5-2B that classifies text, routes requests, and scores content against a rubric with typed answers your application can use directly.

Include several questions in one request to get choices, yes/no probabilities, and scores. Questions about the same state reuse its input prefix, so they cost less together than in separate requests.

Call decision-models/baseten-joey-2b-latest through the System One endpoint with your Baseten API key:

import os
import requests
resp = requests.post(
    "https://inference.baseten.co/v1/systemone",
    headers={
        "Authorization": f"Bearer {os.environ['BASETEN_API_KEY']}",
    },
    json={
        "model": "decision-models/baseten-joey-2b-latest",
        "state": "I was charged twice for my subscription. Please refund the duplicate charge.",
        "questions": {
            "route": {
                "type": "choice",
                "instructions": "Which team should handle this request?",
                "criteria": {
                    "billing": "Billing, payments, and refunds",
                    "technical_support": "Technical troubleshooting",
                },
            },
            "refund": {
                "type": "noul",
                "instructions": "Is the customer asking for a refund?",
            },
            "urgency": {
                "type": "score",
                "instructions": "How urgent is the request?",
                "criteria": [
                    "Routine request",
                    "Needs attention today",
                    "Immediate emergency",
                ],
            },
        },
    }
)
print(resp.text)

For more information, see our docs.

Starting October 28, 2026, Baseten APIs will return 401 Unauthorized instead of 403 Forbidden when a request has missing, invalid, or expired credentials. This includes requests with invalid API keys or expired authentication tokens.

Before the rollout, update any clients, monitoring, or automation that check for 403 to detect authentication failures. These should handle 401 responses instead.

Requests with valid credentials but insufficient permissions will continue to return 403 Forbidden.

For more information, see our docs.

Starting November 5, 2026, Baseten will automatically revoke API keys that haven’t authenticated a request in 90 days. This change is part of our ongoing work to strengthen workspace security and reduce the risk from unused credentials. Keys that remain unused for 90 days will continue to be revoked on an ongoing basis.

Before enforcement begins, review the last-used timestamp for each key on your API keys page. Rotate any keys used by infrequent workloads and update those workloads with the replacement keys. If your team provisions keys through automation, review those workflows as well.

Revoked keys cannot be restored. Requests using a revoked key will fail authentication.

For more information, see our documentation.

eMudhra/emSign, a third-party certificate authority, has reported that some of its signing keys were compromised. Baseten's systems and data are not affected.

We've rebuilt the Baseten (v1.4.1) and Truss CLIs (0.18.36) with fixes to block certificates from affected authorities.

We strongly recommend upgrading to the latest versions of these, and using Baseten CLI moving forward. If you don't upgrade and you're running with outdated certificate bundles on your OS, your traffic can be intercepted and decrypted on an untrusted network. Note that most Windows & MacOS systems have updated bundles and are not affected.

Instructions:

Baseten CLI:

brew update && brew upgrade baseten

Truss:

pip install truss==0.18.36

You can now track three additional GPU metrics in the Metrics tab for your Dedicated Inference deployments:

• GPU temperature: Temperature in degrees Celsius.
• GPU tensor core utilization: Tensor core activity as a percentage, on GPUs with profiling support.
• GPU memory bandwidth utilization: Memory read and write activity as a percentage.

Compare these graphs with request load and latency to investigate performance bottlenecks. Use Customize view in the Metrics tab to show the graphs you need.

For more information, see our docs.

elephant-1 and mammoth-1

elephant-1 and mammoth-1 are now available through Baseten Model APIs. Classify text, route requests, and score content against a rubric. Send one or more questions in a single request to get choices, yes/no probabilities, and scores.

Call the System One endpoint with your Baseten API key:

curl https://inference.baseten.co/v1/systemone \
-H "Authorization: Bearer $BASETEN_API_KEY" \
-H "Content-Type: application/json" \
-d '{
  "model": "decision-models/baseten-elephant-9b-latest",
  "state": "I was charged twice for my subscription. Please refund the duplicate charge.",
  "questions": {
    "route": {
      "type": "choice",
      "instructions": "Which team should handle this request?",
      "criteria": {
        "billing": "Billing, payments, and refunds",
        "technical_support": "Technical troubleshooting"
      }
    }
  }
}'

To use mammoth-1, set model to decision-models/baseten-mammoth-27b-latest.

For more information, see our docs.

Reflection's Beam is now available through Baseten Model APIs as reflection/Beam-1.0. It supports a 1M-token context window, tool calling, and always-on reasoning. Set reasoning_effort from minimal to max to control the reasoning budget.

curl https://inference.baseten.co/v1/chat/completions \
-H "Authorization: Bearer $BASETEN_API_KEY" \
-H "Content-Type: application/json" \
-d '{
  "model": "reflection/Beam-1.0",
  "messages": [{"role": "user", "content": "Can cars fly yet?"}],
  "reasoning_effort": "minimal",
  "max_tokens": 1024
}'

For multi-turn conversations, preserve reasoning_content from prior assistant turns.

For more information, see our docs.

Meta Muse Spark Open and NVIDIA Nemotron 3.5 Super VL are now available through Baseten Model APIs. Send requests through the OpenAI-compatible endpoint with your Baseten API key.

Muse Spark Open

Muse Spark Open accepts text, image, audio, and video inputs and generates text responses. Use it to reason over images, recordings, and video in your application.

curl https://inference.baseten.co/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $BASETEN_API_KEY" \
-d '{
"model": "meta-models/muse-spark-open",
"messages": [{"role": "user", "content": "Implement a Python function that groups events by user ID."}]
}'

Nemotron 3.5 Super VL

Nemotron 3.5 Super VL accepts text and image inputs through Model APIs, with a context window of up to 1,048,576 tokens.

Nemotron 3.5 Super VL is also available for managed LoRA fine-tuning with the Loops SDK. Train with the BF16 or NVFP4 model using a context window of up to 262,144 tokens.

For more information, see the Model API docs and Loops supported models.

Baseten now supports restricting a deployment’s outbound network traffic to an allow list of IP ranges and domains. Add egress_restrictions under runtime in your Truss config to lock down a sensitive model without affecting any other deployment in your organization.

runtime:
  egress_restrictions:
    ip_allow_list:
      - 203.0.113.10/32
    fqdn_allow_list:
      - "*.baseten.co"
      - huggingface.co

Contact us to enable egress restrictions for your organization. For more information, see our docs.

DeepSeek V4.1 Flash Fast is now available through Baseten Model APIs. It serves the same 552B-parameter multimodal model as DeepSeek V4.1 Flash at full precision, on dedicated Fast capacity for higher sustained throughput, and is purpose-built for long-running agentic coding sessions. It has a 1M-token context window, vision, tool calling, and reasoning on by default.

Send requests to deepseek-ai/DeepSeek-V4.1-Flash-Fast through our OpenAI-compatible endpoint with your Baseten API key:

curl https://inference.baseten.co/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $BASETEN_API_KEY" \
  -d '{"model": "deepseek-ai/DeepSeek-V4.1-Flash-Fast", "messages": [{"role": "user", "content": "Read the failing test, find the bug, and send me the patch."}], "reasoning_effort": "high"}'

For more information, see our docs.

You can now rename a model from the terminal with baseten model rename. The model keeps its ID, endpoints, and deployments, so only the name in config.yaml needs updating.

Version 1.1.0 also adds baseten volume sync, in pre-release. Volume syncs can copy a Hugging Face, S3, GCS, Azure, R2, or CoreWeave source into a volume as a durable background job. Authentication with stored keys is supported across providers, and short-lived OIDC is supported to AWS (for S3) and GCP (for GCS). Pass --wait to block until the sync finishes; pressing Ctrl+C stops waiting without cancelling the sync.

We've also updated support for volume listings: baseten volume stat, and baseten volume versions now show when volumes, tags, and versions expire. baseten volume push and baseten volume pull report total transfer time, and concurrent pushes to the same volume fail with a clear message instead of a raw conflict error.

brew upgrade baseten
baseten model rename --model-id <model-id> --new-name my-model-v2
baseten volume sync start --source hf://<organization>/<repository> --dest bdn:<namespace>/<volume>:<tag> --wait

For more information, see our docs.

Starting Thursday, October 1 at 15:00 UTC, newly created Baseten API keys will add a b10_ prefix, making them easier to identify in code, logs, and secret scanners.

Existing keys will continue to work unchanged. The prefix applies only to newly created keys, and no action or key rotation is required.

b10_abcd1234.ABCDEFGHIJKLMNOPQRSTUVWXYZabcdef

If your tooling validates, parses, redacts, or scans API keys based on their current format or length, please update it before October 1.

For more information, see our docs.