THE SOURCE IN CONTEXT
Official Flash checkpoint and local-serving references. Its reasoning effort and chat-template controls matter when attempting evaluation reproduction.
What distinguishes Flash from the full GLM model?
The Flash card describes a newly trained, natively multimodal model with 320B total and 18B active parameters, using sparse and linear attention with mHC. The deployment recipe rounds the total differently at 321B. This page retains the publisher’s wording and the recipe’s number in their own records; neither establishes memory use under a serving workload.
Source: zai-org · pinned document
Which reasoning controls affect reproduction?
reasoning_effort accepts low, high and max. Omitting it, or supplying another value, selects max; the publisher asks benchmark reproductions to retain that setting. clear_thinking is independent and defaults to false. The card recommends explicitly enabling clear_thinking for chat. Keep both settings in an evaluation record rather than inferring one from the other.
Source: zai-org · pinned document
Why are benchmark context budgets different?
HLE with tools uses a 300,000-token managed context and a 163,840-token generation ceiling, with temperature 1.0 and top_p 0.95. DeepSWE instead specifies 400K context, temperature 0.95, top_p 1.0 and a six-hour timeout. These are task-specific harness settings, not one global chat preset or a local speed measurement.
Source: zai-org · pinned document
Where should a deployment begin?
The card links model-specific vLLM, SGLang, TokenSpeed, Transformers and other integration guides. Choose a versioned engine route, then match the actual artifact and hardware profile. A link in the publisher card establishes an intended integration path; this collection has not installed those stacks or verified their performance on a reader’s device.
Source: zai-org · pinned document
What does the source specify?
These declarations belong to the cited document. A planning minimum, model limit and measured runtime allocation are different quantities.
| Total / active parameters as counted by card | 320B / 18B |
|---|---|
| Reasoning effort | low / high / max; default max |
| Chat thinking cleanup | clear_thinking false by default; true recommended for chat |
| HLE-with-tools context / output cap | 300000 / 163840 tokens |
AI-assisted source reading and original editorial synthesis; no independent model execution.