Show HN: Flide: Make your decision models fast

News · Speed · Exact answers · Quick start · Models · Options

Flide is fast inference infrastructure for Jev-style decision models. With exact state forking it is up to 9.4 times faster than vLLM and SGLang, and every answer is the one the model gives the question alone. It serves the Jev /v1/systemone API.

• 🔥 2026-11-09 We addedpplx-decider-v1.1-27b andJEV-35B in Flide 0.1.9, and/v1/systemone now takes several requests in one body.
• 🔥 2026-11-08 We added two decision model families in Flide 0.1.8, includingAutoTrust andClef , and alsodrex-max-v1 .
• 🔥 2026-11-02 We released our paperFlide: Accelerating Decision Inference with Exact State Forking on arXiv.
• 🔥 2026-10-29 We added three decision model families in Flide 0.1.7, includingSynACK Decide ,Wald andXor , and Gemma 4 mixture-of-experts models such asKnowLine-26B-A4B .
• 🔥 2026-10-29 We added two decision model families in Flide 0.1.6, includingDrex andKnowLine , andflide verify checks exact forking on your own model and requests.
• 🔥 2026-10-28 We added six decision model families in Flide 0.1.5, includingJeffan ,JevAny ,Kev ,Lychee ,pplx-decider andStartLux , and images forLumma andmammoth .
• 🔥 2026-10-26 We added five decision model families in Flide 0.1.4, includingcerno ,JADE ,Linne ,mammoth andSieve , and alsoQuyet-1.0-Large .
• 🔥 2026-10-26 We added eight decision model families in Flide 0.1.3, includingAJev ,Bespoke-Nimble ,Darwin ,jiwo ,Lumma ,mogan ,Quyet andSenova , and LoRA adapters now serve as they are.
• 🔥 2026-10-25 We added two decision model families in Flide 0.1.2, includingIntern-Decision andtorchcast , and a question now takes up to 256 options.
• 🔥 2026-10-24 We released Flide 0.1.0 onPyPI , servingQwen3.5 andGemma 4 chat models and theOneJev andOpen-Jev decision models. Welcome to have a try.

With three questions, a generation engine runs three sequences and computes 4 cells of the state again, and multi-item scoring adds 3 delimiters that change the answers. Flide runs one sequence, computes nothing twice, adds no tokens, and gives each question the answer it gets when asked alone.

On one H200 with Qwen3.5-0.8B, Flide is up to 9.4 times faster than both vLLM and SGLang (QuALITY, where it is 14 times faster than vLLM), 3.6 to 4 times faster on RACE and the Decision Index, and serves 2.4 times the throughput with 64 clients. The more questions a state carries, the bigger the gap.

Flide stays within 1.1 points of a float32 reference, the closest of any engine, and flips at most 2.3% of decisions at 0.5. Multi-item scoring moves answers by up to 14.8 points and flips up to 28.9%. With --batch-invariant, a request gets the same bits however it is batched.

See the docs for every model, the API and all options.

uv venv -p 3.12 && source .venv/bin/activate
uv pip install flide torchvision

Any chat model on Qwen3.5 or Gemma 4 serves as it is and answers through its next token.

flide serve --model Qwen/Qwen3.5-4B

OneJev models read their answers their own way, and Flide serves them as they are too.

flide serve --model OmniJev/OneJev1.6-0.8B

An Open-Jev release is an adapter on a Qwen3.5 base. Merge it into a checkpoint once, then serve that.

python -m flide_open_jev.prepare ZefanCai/Open-Jev-2B open-jev-2b
flide serve --model open-jev-2b

flide serve listens on all interfaces at port 8000; pass --host 127.0.0.1 to keep it local.

curl http://localhost:8000/v1/systemone -H "Content-Type: application/json" -d '{
  "state": "Order #1182 arrived with a cracked screen. The customer asks for a refund and attached two photos.",
  "questions": {
    "action": {"type": "choice", "instructions": "What should support do next?",
               "criteria": {"refund": "refund the order", "replace": "send a replacement", "escalate": "ask a human"}},
    "urgent": {"type": "noul", "instructions": "Does this need a reply within the hour?"}
  }
}'

OneJev models also read images. Put <image:1> where the photo belongs in the state and send the photo in media.

IMG=$(base64 -w0 photo.jpg)
curl http://localhost:8000/v1/systemone -H "Content-Type: application/json" -d @- <<EOF
{
  "state": {"ticket": "Order #1182 arrived with a cracked screen. The customer asks for a refund.", "photo": "<image:1>"},
  "media": [{"type": "image", "data": "data:image/jpeg;base64,$IMG"}],
  "questions": {
    "cracked": {"type": "noul", "instructions": "The photo shows a cracked screen"},
    "action": {"type": "choice", "instructions": "What should support do next?",
               "criteria": {"refund": "refund the order", "replace": "send a replacement", "escalate": "ask a human"}}
  }
}
EOF

Every answer carries the probability of each option, ready for a threshold.

The last three columns say how much faster Flide answers one request than the engine its release ships, as the median latency over the requests of each workload on one H200, where DI is the Decision Index. The chat models run against the faster of vLLM and SGLang, and OneJev against qev, the engine it used before Flide. KnowLine 35B runs with --quantize fp8, as its release serves it in FP8. A value marked ~ is the average over the sizes tested.

Apache-2.0. flide/kernels/gdn.py holds a kernel derived from two kernels of flash-linear-attention (MIT); THIRD_PARTY_NOTICES lists it with the other code Flide builds on.

If you use Flide in your research, please cite our paper.

@article{li2026flide,
  title   = {Flide: Accelerating Decision Inference with Exact State Forking},
  author  = {Li, Bobo and Song, Jiajia and Fei, Hao and Hsu, Wynne and Lee, Mong Li},
  journal = {arXiv preprint arXiv:2610.42148},
  year    = {2026}
}