Twenty rivals to TypeSafe's Jev shipped in four weeks. Jev still wins.

At least 20 alternatives to TypeSafe's Jev shipped in roughly four weeks after the model launched on September 15, 2026. They come from Cloudflare, OpenAI, Perplexity, Liquid AI, AWS's Strands Labs and independent developers — and almost every benchmark score in the field is vendor-reported.

Jev is a decision model, not a chatbot. Developers define the allowed answers, and Jev returns a label, a score or a yes/no probability in a single pass, generating no text. TypeSafe prices it at $0.042 per 1M input tokens with free output, 64k tokens of context, closed weights and waitlist access. Its own workflow evals report 67.8% agreement and 0.4 seconds per case.

The stakes are straightforward. If typed decisions (models that answer structured questions — Choice, Score or Noul, a yes/no probability — instead of generating text) replace generative calls for classification-heavy workloads, the cheapest and best-calibrated model wins high-volume pipelines. Four gaps drive the alternative market: no images, no self-hosting, no published size, no fine-tuning.

Takeaway: Jev defined a new API category, and the market rushed in faster than anyone validated the results.

What do the hosted competitors offer?

Cloudflare announced Clef on October 1, 2026. Clef is 27B on Qwen3.8-27B with a vision encoder; Clef-flash is 9B on Qwen3.5-9B. Both are Apache-2.0 on Hugging Face and served on Workers AI at $0.24 and $0.038 per 1M input tokens. Cloudflare calls the API fully Jev-compatible.

In Cloudflare's own run, Clef scored 94.20 BANKING77 F1 versus Jev's 79.74. Jev still wins GPQA Diamond, 78.3 versus 48.0. That split — narrow classification versus hard reasoning — repeats across the entire field.

OpenAI entered with a Decisions API in public beta on October 6, 2026, running only on gpt-6-luna. It costs $0.10 per 1M input tokens, accepts text and base64 images, and uses its own schema rather than Jev's. OpenAI claims roughly 10x faster responses than its Responses API but publishes no benchmarks.

Perplexity's pplx-decider-v1-27b is an Apache-2.0 Qwen3.8-27B fine-tune at $0.02 per 1M input tokens; a 4B flash-v1.2 costs $0.01, the cheapest hosted call in the field. On Perplexity's 11-task panel it scored 85.71% versus Jev's 84.51%, but Jev won BBH, 94.27 versus 82.80.

Liquid AI's d1 family spans hosted models plus open d1-3B (3.1B) and d1-omni-600M (587M), the latter accepting audio clips up to 30 seconds. Liquid reports 48.57 on Decision Index 0.2.1 and 8 ms latency on an RTX 4090 — but does not disclose hosted parameter counts or paid rates.

Smaller entries round out the hosted tier: meraGPT's Decider 1 (September 22, 2026, $0.04 per 1M tokens, Jev-compatible /v1/systemone), Respan's behavior-scoring Span-01 ($0.02, with a free Lite), and Standard Thinking's Standard One in 3B and 8B, both open and hosted.

Takeaway: hosted rivals beat Jev on price and image input, not on the benchmarks Jev itself publicizes.

Which open models can you actually self-host?

The open-weight field runs from 287M to 35B-A3B. Together AI trained Tev1-4B-experimental as a LoRA SFT of Qwen3.5-4B for about $17 in 25 minutes on 37,840 examples — proof that credible decision models need neither a cluster nor a budget. Its weights license is still being finalized, and Together warns its logprobs are not calibrated confidence.

AWS's Strands Labs introduced Strands Decider on October 1, 2026: roughly 1.9B, built on Qwen3.5-2B-Base with a pointer head, Apache-2.0, with 246,678 training rows and the full recipe published. One RTX 3090 trains it in about 11 hours. The v1 checkpoint scored 77.9% on JevBench public.

Independent projects dominate the count. Jared Palmer's Kev (0.8B to 27B, Apache-2.0, TypeSafe SDK works unchanged) fits a 24 GB GPU. Mapika's Decider family ships 0.8B to 35B-A3B; decider-12b tied Jev at 0.730 on JevBench public hard. Zefan Cai's Open-Jev reports 85.28% for the 27B v1.1 on a 231-task JevBench subset. razorback16's OpenJev, built on NVIDIA's DiffusionGemma 26B-A4B, reports 87.4% on JevBench public and offers a free hosted endpoint with 100M tokens per account.

Encoder-based models serve CPU and edge use: Fastino's GLiNER2.5-Decide (340M DeBERTa-v3-large, 167.3 ms p50 on a 48-vCPU Xeon), Convai Innovations' Laya (421M English, 322M multilingual, 32.8 ms p50 on a T4) and wfzyx's Von (395M, 96 ms p50 on 4 vCPUs).

Academic work is present too. InnerJev, from a Fudan, Tongji and King's College London team, distills LLM reasoning into a single-pass decision and introduces JEval, an 11,257-instance eval across 36 datasets. On JEval, InnerJev-27B scored 79.24% versus Jev's 78.38%, with an ECE (expected calibration error, lower is better) of 0.0149 against Jev's 0.0532.

Takeaway: self-hosting a competent decision model now costs between $17 and one consumer GPU.

What does the only third-party run show?

One multi-model comparison exists, published by Respan on October 16, 2026 — and Respan sells a competing model, Span-01. Its Span-01 judge scored 11 models.

Model	Accuracy	Paired flip rate	Injection ASR	ECE
Jev 1.13.0	0.932	0.021	0.063	0.045
Tev1 4B	0.901	0.052	0.042	0.031
Kev 4B	0.833	0.060	0.078	0.070
Decider 2B v10	0.747	0.128	0.208	0.074
Laya	0.542	0.298	0.396	0.092
Von 1.2	0.422	0.242	0.297	0.071

Jev led accuracy and consistency. Together's Tev1 beat it on injection resistance (0.042 versus 0.063 attack success rate) and calibration. Von 1.2 scored the lowest accuracy in the table at 0.422, against Jev's 0.932, and its own repo notes its calibration is tuned to JevBench and can mislead out of domain.

The caveat writes itself: the suites differ — JevBench, Decision Index, typed-decisions, Respan's behavior benchmark — and scores across them are not comparable. Almost every headline number in this market is self-reported.

Takeaway: one conflicted third party is the entire independent evidence base, and it still favors Jev.

Which alternative fits which need?

• Images: Clef, pplx-decider, Standard One, or the DiffusionGemma OpenJev.
• Drop-in Jev swap: Clef, Kev, Mapika Decider, meraGPT or Laya expose /v1/systemone.
• CPU or edge: Von, Laya, GLiNER2.5-Decide, Liquid's d1-omni-600M.
• Train your own: Strands Decider and Tev1 publish full recipes.
• Cheapest hosted call: Perplexity's flash-v1.2 at $0.01 per 1M input tokens.

Whatever the choice, test on your own labeled data first. Decision models can return a valid answer that is still wrong. With 20 rivals in four weeks and no independent benchmark beyond Respan's, the category's next test is whether any of them survive contact with production traffic.

By James Calloway