Skip to main content

GLM-4.7 Flash

Large MoE assistant model served tensor-parallel across four Intel Arc Pro B70s.

Overview​

  • model_name: hf-zai-org-GLM-4.7-Flash (HuggingFace import, currently free)
  • Source: zai-org/GLM-4.7-Flash
  • Architecture: Glm4MoeLiteForCausalLM — Mixture-of-Experts (64 routed experts) with Multi-head Latent Attention (MLA)
  • Served on: GPU via vLLM, tensor-parallel across 4× Intel Arc Pro B70 (62 GB of weights, ~15.5 GB per rank)
  • Status: ready (verified end-to-end through the public inference route, 2026-08-03)

When to use​

✅ Long-form assistant work where you want frontier-class quality: analysis, structured reasoning, and multilingual generation. The model thinks through the request before answering.

❌ Latency-sensitive turns — this is the largest model in the catalog and answers deliberately. For fast replies use Qwen 2.5 7B; for code use Qwen3 Coder 30B.

Live specs​

curl "https://api.colabhive.com/api/builder/v1/endpoints?visibility=public&search=GLM-4.7"

Quick start​

curl -X POST "https://api.colabhive.com/api/builder/v1/endpoints/<endpoint_id>/infer" \
  -H "X-API-Key: $COLABHIVE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": {"messages": [{"role": "user", "content": "Compare optimistic and pessimistic concurrency control, with a worked example."}], "max_tokens": 800}}'
First call after idle

A cold start of a 62 GB tensor-parallel model takes several minutes while the weights are read and the KV cache is profiled. The request is queued, not rejected — it completes once the replica is warm.