48 releases · 0 this month
- Added Kimi K3, GLM 5.3 Flash, Inkling, Reflection AI Beam, and MiMo V2.6 Pro to the Dedicated model catalog. Documented images in tool messages for enabled Gemma 4 31B, Kimi K2.7 Code, and Kimi K3 Dedicated endpoints in Image Inputs. Added Kimi K3 and GLM 5.3 Flash reasoning controls to Reasoning and OpenAI compatibility.
Updated the Service Tiers guide and Chat Completions API reference to match the current service-tier behavior. Responses report the effective tier in service_tier. The documentation now also explains how auto can fall back to flex and how priority can retry at the request's standard priority.
Temporary increase to Qwen 3.8 27B total token rate limit The Developer tier total token rate limit for qwen-3.8-27b is temporarily increased from 450K to 750K tokens per minute (5× the uncached limit, up from the standard 3×). The uncached limit remains 150K tokens per minute. See Rate Limits for current limits.
Qwen 3.8 27B available on public endpoints qwen-3.8-27b is now available on Cerebras public endpoints. Qwen 3.8 27B is a 27-billion-parameter dense multimodal model for agentic coding, tool use, research, and long-running workflows.
Flowise integration moved to legacy status Flowise has archived its repositories and ends official support on August 31, 2026. We removed Flowise from the supported integrations directory and navigation. The legacy Flowise guide remains available for existing self-hosted deployments, but we do not recommend starting new production projects with Flowise.
GLM 4.7 reasoning_logprobs default will not ship before deprecation zai-glm-4.7 is scheduled for deprecation on August 17, 2026 and will keep its existing default logprobs response shape until then. Requests that explicitly set the X-Cerebras-Version-Patch: 2 header still receive the separated reasoning_logprobs field.
Image dimension limit for image inputs Image inputs now enforce a maximum width and height of 15,000 pixels per side. Requests with larger images return 413 Content Too Large with the message "Image dimensions {width}x{height} exceed the 15000px per-side limit". Existing payload-size, image-count, and decompressed-memory limits are unchanged.
Predicted Outputs limited to dedicated endpoints Predicted Outputs is available by request on dedicated endpoints. To request it for your endpoint, contact us or reach out to your account representative. You can also try Predicted Outputs with gpt-oss-120b in Public Preview using a Free Trial or Developer account.
Image limit increased for Developer and Enterprise accounts Developer and Enterprise accounts using gemma-4-31b on public endpoints can now include up to 10 images in a single request, increased from 5. Free Trial accounts remain limited to 2 images per request.
API Version 2 Default Rollout As of July 22, 2026, API version 2 is the default. Version 2 includes: Stricter schema validation for structured outputs Stricter validation for multi-turn tool-calling messages A separate reasoning_logprobs response field for reasoning-token logprobs Corrected Unicode handling in logprobs Z.ai GLM 4.7 is scheduled for...
Free Trial credits for new accounts New accounts now receive \$5 in free credits after adding a verified payment method. Credits work across all models on public endpoints and expire 30 days after Cerebras grants them. You aren't charged unless you purchase additional credits.
Gemma 4 31B available across coding integrations gemma-4-31b is now a selectable model in our Cline, Kilo Code, opencode, and VS Code integration guides. Gemma 4 31B adds multimodal support to coding workflows , you can pass screenshots, diagrams, and design assets alongside text prompts.
Gemma 4 31B and Image Input Support Gemma 4 31B (gemma-4-31b) is now available in preview. It's the first model on Cerebras to support image input. Send screenshots, documents, diagrams, and design assets alongside text using the standard Chat Completions API. Image Inputs are now available in Public Preview through the Chat Completions API.
New dedicated models: StepFun Step 3.5 Flash and Step 3.7 Flash StepFun's Step 3.5 Flash and Step 3.7 Flash are now supported on Dedicated Endpoints.
Projects is now generally available Projects is out of private preview and available to all organizations. Use Projects to group API keys, set per-project rate limits, segment usage analytics, and manage member access across isolated workspaces. Every organization starts with a Default Project, so existing setups are unchanged.
New dedicated models: GLM 5, GLM 5.1, and Kimi K2.6 Z.AI's GLM 5 and GLM 5.1 and Moonshot AI's Kimi K2.6 are now supported on Dedicated Endpoints.
Validation Errors Now Return 400 Instead of 422 API requests that fail validation now return HTTP 400 Bad Request instead of 422 Unprocessable Entity, aligning the HTTP status code with standard behavior across major inference providers.
New parameter: prompt_cache_key /v1/chat/completions and /v1/completions now accept an optional prompt_cache_key parameter. Requests that share the same key are routed together so they reuse the same prompt cache, increasing cache hits and reducing time to first token.
Payload Optimization: msgpack and gzip /v1/chat/completions and /v1/completions now accept request bodies encoded with msgpack (Content-Type: application/vnd.msgpack) and/or compressed with gzip (Content-Encoding: gzip). This can significantly reduce payload size and improve TTFT for requests with long prompts.
New Sampling Parameters The following parameters are now available on all models: frequency_penalty – Reduces repetition by penalizing tokens based on how often they have appeared so far presence_penalty – Reduces repetition by penalizing tokens that have appeared at least once logit_bias – Adjusts the likelihood of specific tokens appearing in the response
Streaming: Server-Sent Events (SSE) Streaming responses for Chat Completions (/v1/chat/completions) and Completions (/v1/completions) now emit a standard data: [DONE] SSE event at the end of the stream, following the SSE standard (see the living specification). The [DONE] marker is guaranteed to appear exactly once and is always the last event in the stream.
Rolling out selective 4-bit weight-only quantization for supported models, targeting non-sensitive layers while preserving full precision for sensitive layers, in line with industry standards. Learn more in Supported Models.
developer Message Role (gpt-oss-120b) The Chat Completions API now supports the developer message role for gpt-oss-120b, aligning with the OpenAI API message format. The developer role is functionally equivalent to system , the system role remains supported for backwards compatibility. Other models do not support the developer role.
Logprob Consistency Logprobs are now computed from the model's raw output, before applying temperature scaling. Previously, logprob values would vary with temperature , they now remain consistent regardless of the temperature used. This aligns with standard behavior across major inference providers and vLLM. Sampled outputs are unaffected.
Deprecating disable_reasoning parameter for zai-glm-4.7 The disable_reasoning parameter is deprecated and will no longer be supported after July 21, 2026. Migrate to reasoning_effort="none" to disable reasoning on zai-glm-4.7. See the Reasoning guide and Chat Completions API reference for details.
reasoning_effort="none" support for GLM 4.7 zai-glm-4.7 now accepts the reasoning_effort parameter. Reasoning is enabled by default , set reasoning_effort="none" to disable it for simple tasks where verbose reasoning blocks are unnecessary. The disable_reasoning parameter continues to be supported but will be deprecated soon.
Monitor your dedicated inference endpoints with the new Metrics API, providing Prometheus-compatible metrics for requests, tokens, latency percentiles, and endpoint health. Metrics are aggregated at the minute level and integrate directly with Prometheus. Learn more in our Metrics guide.
API Version 2 Available for Testing API version 2 is now available for testing via the X-Cerebras-Version-Patch header. This version introduces stricter validation and breaking changes.
Prioritize requests to the API to balance latency sensitivity and resource allocation across your workloads with the new Service Tiers feature. Learn more in our Service Tiers guide.
Constrained Decoding Generally Available Constrained decoding has moved to General Availability (GA) with expanded capabilities. Supported models Structured Outputs and Tool Calling with strict: true are now available on the following models: gpt-oss-120b llama-3.1-8b qwen-3-235b-a22b-instruct-2507 qwen-3-32b zai-glm-4.6 zai-glm-4.7 New capabilities...
Process large-scale inference workloads asynchronously with the new Batch API and Files API. Submit up to 50,000 requests in a single batch job and get 50% cost savings. Learn more in our Batch guide.
Added support for parallel tool calling, allowing models to request multiple tool invocations simultaneously in a single response. This reduces latency when handling queries that require multiple independent operations. Learn more in our Tool Calling guide.
Streaming Support Streaming is now supported for all models, including with structured output requests (response_format, tools). Reasoning Format Controls Added reasoning_format parameter for reasoning models, giving you control over how reasoning text appears in responses.
Added support for prompt caching, which automatically stores and reuses prompt prefixes to reduce latency and improve time to first token for similar requests. Learn more in our Prompt Caching guide.
We now support Predicted Outputs for faster generation when large portions of the output are known in advance. Provide a draft of the expected response via the prediction field, and the model will efficiently reuse unchanged tokens while regenerating only those that differ. Ideal for code editing, document revisions, and template-based responses.
Deprecated qwen-3-coder-480b. We recommend migrating to Z.ai GLM 4.7.
Added support for retrieving log probabilities via the logprobs parameter for the Completions endpoint. Set to an integer (0-20) to return the most likely alternative tokens at each position, or null to disable.
We're gradually introducing multi-token streaming across our models. Instead of sending tokens individually, we'll now deliver them in batches through 200 evenly-spaced events per second. This change eliminates the artificial delays that occurred with single-token streaming.
Updated support for OpenAI GPT-OSS (gpt-oss-120b) Tool calling with strict: true (constrained decoding) Response format with json_schema when using strict=true Response format with json_object when using strict=true tool_choice: none Updated JSON Schema limitations: Nested levels increases from 5 to 10 Max properties increases from 100 to 500 Enabled...
Fast inference on wafer-scale chips.
| Release-notes page | Site | Feed |
|---|---|---|
| Release Notes training-docs.cerebras.ai | training-docs.cerebras.ai | None |
| Release Notes sdk.cerebras.ai | sdk.cerebras.ai | None |
| Software Release Notes training-api.cerebras.ai | training-api.cerebras.ai | None |
Releasesby Glyph