fix: add Cerebras models zai-glm-4.7

[github-actions]

fix: add Cerebras models zai-glm-4.7

Closes #590

Source issue: #590

Summary

Field	Value
Provider	cerebras
Primary model	zai-glm-4.7
Changed models	zai-glm-4.7
Added models	zai-glm-4.7
Updated models	None
Verification sources	1 2 3 4

Verified metadata

Model	Display name	Parent	Providers	Format	Flavor	Token limits	Pricing	Lifecycle
zai-glm-4.7	Z.ai GLM 4.7		cerebras	openai	chat	input=131072, output=40960	in/out=2.25/2.75 per 1M	reasoning=true

Verification notes

Verification

Official sources consulted

1. Cerebras Models Overview — https://inference-docs.cerebras.ai/models/overview

   • Verified: both model IDs exist as Preview models on Cerebras inference
   • Verified: model parameter counts (355B for GLM 4.7, 235B for Qwen 3)

2. Cerebras Pricing — https://cerebras.ai/pricing

   • Verified: zai-glm-4.7 input $2.25/M, output $2.75/M
   • Verified: qwen-3-235b-a22b-instruct-2507 input $0.60/M, output $1.20/M

3. Cerebras Z.ai GLM 4.7 Model Page — https://inference-docs.cerebras.ai/models/zai-glm-47

   • Verified: context window 131k tokens (paid), max output 40k tokens
   • Verified: reasoning model (reasoning enabled by default, reasoning_effort parameter)
   • Verified: text-only input/output (not multimodal)
   • Verified: supports streaming, structured outputs, tool calling, prompt caching
   • No deprecation date listed

4. Cerebras Qwen 3 235B Model Page — https://inference-docs.cerebras.ai/models/qwen-3-235b-2507

   • Verified: context window 131k tokens (paid), max output 40k tokens (paid)
   • Verified: non-thinking mode only ("This model supports only non-thinking mode")
   • Verified: deprecation date May 27, 2026
   • Verified: text-only input/output (not multimodal)
   • Verified: supports streaming, structured outputs, tool calling, prompt caching

sync_models (LiteLLM) cross-check

Neither cerebras/zai-glm-4.7 nor cerebras/qwen-3-235b-a22b-instruct-2507 exists in the LiteLLM model_prices_and_context_window_backup.json catalog. All proposed values are sourced directly from official Cerebras documentation. No deviations to report because sync_models has no entries for comparison.

Token limit interpretation

Cerebras docs report token limits as approximate values ("131k", "40k"). Based on the existing catalog entry for gpt-oss-120b on Cerebras (max_input_tokens: 131072, max_output_tokens: 32768) and standard LLM conventions where "32k" = 32768 (32 * 1024), the values are interpreted as:

• "131k" input = 131072 tokens (128 * 1024, matching existing Cerebras models)
• "40k" output = 40960 tokens (40 * 1024, following the same convention)

Fields not published or not applicable

• parent: No stable alias exists for either model on Cerebras; not applicable
• input_cache_read_cost_per_mil_tokens / input_cache_write_cost_per_mil_tokens: Cerebras docs mention prompt caching support but do not publish separate cache pricing; omitted
• supported_regions: Not applicable (Cerebras is not a Vertex provider)
• locations: Not applicable (Cerebras models do not require explicit location metadata)
• multimodal: Both models are text-only; omitted (defaults to falsy)

Deprecation note

qwen-3-235b-a22b-instruct-2507 has deprecation_date 2026-05-27, which is within the resolver's 90-day deprecation window. The resolver will automatically skip this model as near-deprecation and proceed with only zai-glm-4.7. This is expected and correct behavior.

sync_models vs proposed update

sync_models cross-check found differences. Official provider verification was used for the applied values, and sync_models discrepancies are listed below for review.

Model	Field	Proposed update	sync_models	sync_models source models
zai-glm-4.7	max_input_tokens	131072	128000	cerebras/zai-glm-4.7
zai-glm-4.7	max_output_tokens	40960	128000	cerebras/zai-glm-4.7

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ai-proxy	Ready	Preview, Comment	May 21, 2026 8:45pm