Models

OpenJev

OpenJev turns DiffusionGemma 26B-A4B, a discrete diffusion language model, into a System One model. It writes the answer template onto the model's canvas, leaves only the answer slots as noise and reads a probability distribution for every slot from a single denoising step. There is no generation and no parsing, so an answer can never fall outside your schema.

Model ids

The response's model field always names the exact version that answered.

OpenJev 0.1 reads its answers through vLLM's structured reads for DiffusionGemma, from pull request #57250. That pull request merged upstream on 2026-09-22, so OpenJev now runs on upstream vLLM instead of a fork. When a new version ships, openjev-latest moves to it and openjev-0.1 stays available for pinning.

Since 2026-10-06 the server (OpenJev 0.6.0) runs upstream vLLM main at a3e0243b, which includes every DiffusionGemma fix merged after #57250, among them a sampler kernel rewrite and a fix for an attention mask that froze under CUDA graph replay. The model and the id stay openjev-0.1. On 2,501 labeled text questions it is as accurate as before (76.7% against 77.0%, within run-to-run noise), but an individual probability can move by a few points.

Since 2026-10-12 every model runs OpenJev 0.7.0, which adds POST /v1/decisions, OpenAI's Decisions API, for all of them. The answers to /v1/systemone don't change.

Laya

Laya is by Nandakishor M at Convai Innovations. It is a small encoder model: it reads each question once, with the state, and scores every option in one forward pass. Codiv serves its laya-typed-decisions checkpoint, which the authors fine-tuned on the typed-decisions workflows (invoice processing, security incidents, customer service and agent traces). For benchmarks and known limits, see the Laya repository.

Verdict

Verdict is by Heman10x. It is a ModernBERT-base encoder with a GLiClass classification head, trained with proper scoring rules (RLCD) and calibrated with a temperature for each option count. Codiv serves the rlcd-modernbert-151m checkpoint with Verdict's v1.4 inference engine; for the same input the probabilities match that engine's own output. For benchmarks and known limits, see the Verdict repository.

Verdict can also answer "insufficient evidence". Codiv drops that probability and scales the others to sum to 1, as the System One answer shapes require. Verdict ignores a noul's criteria.

What differs from OpenJev

• Text only. images, steps above 1, samples above 1, think and sequential return 400.
• Each question is read with its own copy of the state, so usage.input_tokens counts the state once per question.
• They are small and fast: a request with 16 questions takes about 10 ms of model time on Laya and 7 ms on Verdict.

Both models are served by OpenJev's encoder backends. The credit for the models belongs to their authors.

CLM

CLM (Contrastive Language Model) is by Contrastive-LM. It puts two small projection heads, a state head and an action head, on top of a frozen Qwen3-8B. The state (with the question appended) and every option are embedded separately, and an answer is the softmax over each option's scaled similarity to the state. Codiv serves the CLM-v0.1-8B heads with Qwen3-8B in FP8. For benchmarks and known limits, see the CLM repository.

Score questions can ignore the state (CLM issue #3); choice and noul questions follow it. Recently seen texts are cached, and usage.input_tokens counts only the texts that had to be embedded, so a repeated request can report 0.

JevK5

JevK5 is by Alibi Serikbay. It is Qwen3.5-4B with a LoRA distilled from Qwen3.6-27B and GPT-6 Luna, merged. Each question becomes a prompt with its options lettered A to P, and the answer is a softmax over those letters' next-token logits under one calibration temperature (the readout is SemIf's). Codiv serves the JevK5 v0.3 weights in bf16. On JevBench's 231 public items it reads the same prompts as the author's own published v0.3 run and gives the same top answer on 230; the other is an exact 50/50 tie. v0.3 replaced v0.2 on 2026-10-12, and jevk5-0.2 is no longer served. For benchmarks and known limits, see the JevK5 repository.

Served from a second GPU box

CLM and JevK5 run on their own GPU box, separate from the one that serves OpenJev, Laya and Verdict. Either box can go down without affecting the other's models; while it is down, its models return 529. They share the differences from OpenJev listed above, and each question reads the state again. The credit for the models belongs to their authors.

Clef-flash

Clef-flash is by Cloudflare. It is Qwen3.5-9B with its vision encoder and a small transformer head that reads the final hidden states and scores every option of every question in one forward pass, so all of a request's questions are answered together. Codiv serves alpha-x-ai's NVFP4/FP8 quantization, which gives the same top answer as Cloudflare's bf16 release on 98.9% of its authors' 1,358 test questions. On Codiv's own checks it gives the same top answer as that checkpoint's runtime on every question, with and without images. For benchmarks, see the Clef-flash model card.

It runs on the same GPU as OpenJev. A request with three questions takes about 30 ms of model time, 35 ms with an image. steps, samples, think and sequential return 400. The credit for the model belongs to Cloudflare.

Coming next

OpenJev 1.0, a new model of our own. Batch mode for large jobs, and more open System One models from the community.