tilde-8b-a1b

Want to run it? Get the GGUF files at resident-labs/tilde-8b-a1b-GGUF and serve them with tilde.cpp. This repository holds the original trained weights.

Tilde is a decision model. Give it a state (a document, a chat, a web page, a tool result, an agent's trajectory) and any number of typed questions about it. It reads the state once and answers every question in a single forward pass, each with a calibrated probability distribution. No text is generated, so there is nothing to parse.

Question type	Answer
choice	a probability for each of up to 255 named options
noul	the probability that a yes/no statement about the state is true
score	a distribution over 2 to 10 described levels

Classify, gate, triage, rank and route, on one 24 GB GPU, with up to 32k tokens of state.

How it reads

Most language models are causal: each token sees only what came before it. A causal reader has already worked through the whole state before it reaches the questions, so it reads without knowing what it is looking for.

Tilde reads the state and the questions together. Its Ling-3.0 backbone mixes 18 linear-attention layers with 6 full-attention layers. The linear layers stay causal; the 6 full-attention layers are bidirectional, so every token (in the state, the questions and the options) can attend to every other token in the request.

We think this is the right design for a decision model:

• The state is read with the questions in view. The parts of a long document, page or trajectory that matter for the questions are picked out where they appear, instead of the state being encoded blind and searched afterwards.
• Options are weighed against each other. Every option sees the others, so the model compares them jointly rather than scoring each one in isolation.
• Many questions share one reading. All questions get full access to the state in the same forward pass, so adding questions is nearly free.

In our ablations, reading bidirectionally improved both accuracy and calibration over a causal reader.

A model pretrained causally does not know how to use look-ahead, so changing the attention mask alone helps little. Before fine-tuning, Tilde gets a short masked next-token prediction (MNTP) warm-up that teaches the backbone to use context on both sides. The recipe follows:

• LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders (MNTP)
• BidirLM: From Text to Omnimodal Bidirectional Encoders by Adapting and Composing Causal LLMs (bidirectional adaptation with an MNTP warm-up)
• dQwen3.5: Hybrid-Attention Diffusion Language Models (making only a hybrid model's full-attention layers bidirectional while its recurrent layers stay causal)

Because the state is encoded in light of its questions, send every question about a state in one request: that is both faster and how the model was trained.

Results

On the Decision Index 0.2.1, Tilde scores 56.3 with 1.3B parameters active per token: level with 27B models, and 6.3 points above the best model under 25B.

Decision Index 0.2.1 table

Model	Active params	Decision Index	Knowledge	Language	Retrieval	Tools	Arts
Jev (hosted)	–	57.9
Surogate Rune 26B-A4B	4B	57.4
Decider · Gemma-4-31B	31B	57.3
pplx-decider 27B	27.8B	56.4
Tilde 8B-A1B	1.3B	56.3	38.7	60.6	57.8	79.2	45.9
Winnow 12B	12B	50.0	33.8	56.0	54.0	71.0	30.0
d1-3B	3.1B	48.6	23.8	56.4	52.8	74.5	36.3
Decider 35B-A3B	3B	47.1	31.8	55.5	54.7	56.5	32.6
JPT 9B	9.7B	46.9	31.7	56.7	44.6	67.0	28.6

Tilde was scored with the official Decision Index kit, edition 0.2.1 (not a leaderboard submission). Other rows are from the public leaderboard snapshot of 2026-09-28, except d1-3B, which is self-reported on its model card.

Head to head with Jev

Full numbers

Benchmark	Tilde	Jev
CLadder	89.0	72.6
HoVer	86.7	72.9
BANKING77	89.2	79.7
FinEntity	93.1	87.0
CLINC150 + OOS	94.8	89.3
ContractNLI	76.5	71.7
GSM8K	84.4	79.9
RAGTruth	75.4	76.5
HellaSwag	92.9	94.5
BFCL	93.5	95.8

Jev is jev-1.13.0, queried through its API on identical questions and scored by the same code.

Quickstart

Run Tilde with tilde.cpp, our llama.cpp fork with the Tilde head and a /v1/systemone server. It runs on NVIDIA, AMD, Apple Silicon and Intel GPUs, or on a CPU.

# build tilde.cpp (see its README), then:
tilde-server        # downloads the GGUF that fits your hardware and serves http://127.0.0.1:8080

The quantized files are in resident-labs/tilde-8b-a1b-GGUF.

Tilde speaks the /v1/systemone protocol:

curl -s http://127.0.0.1:8080/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "Charged twice for my March invoice. Need the refund before Friday.",
  "questions": {
    "spam":    {"type": "noul", "instructions": "Is this spam?"},
    "team":    {"type": "choice", "instructions": "Which team should handle it?",
                "criteria": {"billing": null, "support": null, "sales": null}},
    "urgency": {"type": "score", "instructions": "How urgent is it?",
                "criteria": ["not urgent", "low", "medium", "high", "critical"]}
  }}'

Files in this repository

These are the trained weights in their native form, for converting or studying, not for running directly: dense_delta.safetensors, routed_experts.safetensors, joint_head.safetensors, markers.safetensors, the tokenizer and report.json. They are applied on top of inclusionAI/Ling-3.0-tiny-base; tilde.cpp's tilde/convert_tilde_to_gguf.py turns the base plus these files into the GGUFs. SHA256SUMS lists every file's hash.

Limitations

Tilde judges what is in the state; it does not look facts up, and multi-step arithmetic is answered in one pass. Put the evidence in the state and tune thresholds on your own data for high-stakes use.

---

_{Built by Resident: small, open models that live on your machine.}