GLiDE typed decisions

Serve Fastino GLiDE decision checkpoints through the typed decision API

GLiDE is a family of Fastino decision models. Given an input and typed questions, a GLiDE checkpoint returns a probability for every option of every question from one forward pass per question. It reads a 255-way decision readout and generates no text. Antfly serves GLiDE checkpoints natively through /ai/v1/decisions using the model-independent typed decision contract. They appear under deciders in model discovery.

Supported checkpoints#

CheckpointBackbonePrompt lengthLonger prompts
fastino/GLiDE-0.4BModernBERT (Ettin) encoder, bidirectional7,999 tokensKeeps the first 1,024 tokens and the end (question and options)
fastino/GLiDE-0.8BQwen3.5 text backbone, non-causal full attention40,000 tokensRejected with HTTP 413
fastino/GLiDE-2BQwen3.5 text backbone (wider), non-causal full attention40,000 tokensRejected with HTTP 413

Antfly recognizes a GLiDE directory from its decision_config.json (format 1, prompt format quick-empty-think-medium-v1) and readout.safetensors. The backbone in config.json must match the declared backbone. Download a checkpoint into the models directory; no conversion is needed:

hf download fastino/GLiDE-0.4B --local-dir ./models/GLiDE-0.4B
antfly inference decisions ./models --request decision.json --backend metal

Text-only bf16 safetensors checkpoints are supported. NVFP4 checkpoints, looped variants and image inputs are rejected. Metal is the qualified backend for every checkpoint. GLiDE-0.4B also runs on the native CPU backend; GLiDE-0.8B and GLiDE-2B are Metal-only, and loading them on another backend fails instead of serving them. The decoder checkpoints run their bf16 projections directly from the mapped safetensors, so loading one reserves its weight size once (3.8 GB for GLiDE-2B) plus a small allowance for prepared tensors.

GLiDE weights are licensed under the Fastino Community License, not Apache 2.0. Commercial use requires displaying "Built with Fastino GLiDE", and production commercial use by an entity with more than US$10 million in annual revenue requires a separate commercial agreement with Fastino. Read the checkpoint's LICENSE before deploying.

Request mapping#

GLiDE prompts are built exactly as the reference glide package builds them, with one prompt per (input, question). Prompt text and token ids are byte-identical to the reference. Public questions map to GLiDE options as follows:

Public questionGLiDE options
choicechoices in order. (A) value: description; a choice without a description renders (A) value.
predicate(A) true: Yes, (B) false: No. probability is p(true).
scorelevels in order as options 0..n-1. The level text is label: description when both are given and differ, otherwise the label. A label that only repeats its index ("0", "1", ...) shows just the description.

The answer distribution is a softmax over the question's options of the readout logits, scaled by the checkpoint temperature. Answers use the typed decision response: choice and its probabilities, the expected zero-based score and labeled probabilities, and the predicate's true probability. confidence is the normalized inverse entropy of the distribution.

Predicates always use the default option text true: Yes and false: No. The reference's own request format can replace that text (for example "criteria": {"true": "Yes, there is a compliance risk", "false": "No compliance risk"}). The public contract has no field for it, so put that wording into the predicate's instructions. The prompt then differs from a reference request that uses custom option text, and so do the probabilities. The reference format also accepts a JSON value as the state; to reproduce its prompt, send Python's json.dumps(state, ensure_ascii=False) rendering as the input string. It also accepts instructions given as objects. Embedding options, multi_choice, choice examples, model_identity and long_document windowing are rejected for GLiDE.

Batching and limits#

All prompts of a request (inputs × questions) are tokenized before execution, sorted by length and run in right-padded micro-batches of at most 16,384 padded tokens (the reference rule). Every micro-batch is admitted against the model's memory limits before it runs and is split further when it does not fit. Token usage counts every prompt's tokens, including the special tokens and after truncation.

The public limits still apply: 128 inputs, 64 questions and 2–64 choices or levels per request. As for Laya, a configured per-item input token limit (max_input_tokens_per_item) applies to each rendered prompt, which includes the question's instructions and options, after the 0.4B's head and tail truncation.

The decoder checkpoints accept prompts up to 40,000 tokens, but each prompt's planned workspace must also fit the server's scratch budget. The default budget is an eighth of the automatic memory envelope: 768 MiB on a 16 GB Mac and at most 1.5 GiB on larger machines. Under the default budget, a single prompt is admitted up to about these lengths:

Checkpoint768 MiB (16 GB Mac)1.5 GiB
GLiDE-0.8B6,500 tokens11,600 tokens
GLiDE-2B4,000 tokens7,700 tokens

Longer prompts return 400 MODEL_RESOURCE_LIMIT. Raise the budget with antfly inference run --scratch-budget-mb <MiB>; a full 40,000-token prompt needs about 8 GiB (0.8B) or 11 GiB (2B). The live-memory check still applies, so the machine must also have that memory free.

GLiDE-2B's resident weights also occupy most of the GPU memory domain on a 16 GB Mac (4.8 GB by default), which leaves about 0.8 GB for a forward's workspace. Longer GLiDE-2B prompts there also need --backend-budget-mb, for example --backend-budget-mb 6144 --scratch-budget-mb 2048 for a 5,000-token prompt.

Validation#

Prompt text and token ids are compared byte for byte with captures from the reference package. The captures cover all three checkpoints, inputs up to 5,129 tokens, truncated encoder prompts and whitespace runs. For eight reference requests translated to the public contract, Metal answers match the reference's CPU FP32 probabilities to within 1e-6 (0.4B), 7e-6 (0.8B) and 6e-6 (2B), with the same top answer for every question. The 5,129-token GLiDE-2B request was run with --backend-budget-mb 6144 --scratch-budget-mb 2048; the default 16 GB Mac budgets reject it.