2026年20款Jev替代方案：挑戰TypeSafe的決策模型
TypeSafe launched Jev on September 15, 2026.It is a decision model, not a chatbot.You define the allowed answers, and Jev returns a label, a score or a yes/no probability.In about 4 weeks, at least 20 Jev alternatives shipped.
They come from Cloudflare, OpenAI, Perplexity, Liquid AI, AWS’s Strands Labs and independent developers.This guide compares each one on size, license, hardware, pricing, Jev API compatibility and published benchmarks.TL;DR Size range: 287M (GLiNER2.5-multi-Decide) to 35B-A3B (Mapika Decider).
Jev’s size is not disclosed.Runs on: Encoders like Von and Laya run on CPU.4B adapters fit a 24 GB GPU or a 32 GB Mac.Performance: Most open models trail Jev on hard reasoning.Several match or beat it on narrow classification.Best: Cloudflare Clef scores 94.20 BANKING77 F1 versus Jev’s 79.
74, per Cloudflare’s own run.Worst: Von 1.2 scores 0.422 accuracy in Respan’s run, against Jev’s 0.932.What is Jev, and why look for alternatives?Jev is TypeSafe’s first ‘System One’ model.You send a state, such as a ticket or log line.You also send typed questions: Choice, Score or Noul (yes/no).
Jev scores every question in one parallel pass and generates no text.TypeSafe’s docs list $0.042 per 1M input tokens, with free output.Context is 64k tokens per request.Input is text only.Weights are closed, and access runs through a waitlist.TypeSafe’s workflow evals put Jev at 67.
8% agreement and 0.4 s per case.That leaves 4 gaps developers keep raising: no images, no self-hosting, no published size and no fine-tuning.The 20 models below fill some or all of them.How should you read these benchmarks?Most figures below are self-reported.
The suites differ: JevBench, Decision Index, typed-decisions and Respan’s behavior benchmark.Scores across suites are not comparable.The one multi-model third-party table comes from Respan, which sells a competing model.Hosted decision APIs 1.
Cloudflare Clef and Clef-flash 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.They accept text, JSON and up to 4 images per request.Workers AI pricing is $0.
24 (Clef) and $0.038 (Clef-flash) per 1M input tokens.Cloudflare calls the API fully Jev-compatible.In its own run, Clef scored 94.20 BANKING77 F1 versus Jev’s 79.74.Jev still wins GPQA Diamond, 78.3 versus 48.0.2.Liquid AI d1 Liquid AI’s d1 family has hosted d1 and d1-ultra, plus 2 open models.
d1-3B (3.1B) runs on LFM2.5-VL-3B.d1-omni-600M (587M) adds audio clips up to 30 s.The open weights use the LFM Open License.Liquid’s guide calls d1 through TypeSafe’s own SDK with a different base URL.Liquid reports d1-3B at 48.57 on Decision Index 0.2.1.It also reports 8 ms latency on an RTX 4090.
Hosted parameter counts and paid rates are not disclosed.3.OpenAI Decisions API OpenAI released the Decisions API in public beta on October 6, 2026.It runs only on gpt-6-luna.It accepts text plus base64 images.Question types are predicate, choice and score.Pricing is $0.
10 per 1M input tokens, with output not billed.The schema is OpenAI’s own, so it is not Jev-compatible.OpenAI claims about 10x faster responses than the Responses API.No benchmarks are published.4.Perplexity pplx-decider Perplexity’s pplx-decider-v1-27b is a Qwen3.8-27B fine-tune under Apache-2.0.
It reads text, JSON and up to 8 images per question.The API allows 128 questions and 255 options per request.The current pricing is $0.02 per 1M input tokens for v1.2 (27B).A 4B flash-v1.2 costs $0.01.On Perplexity’s 11-task panel, it scored 85.71% versus Jev’s 84.51%.Jev still won BBH, 94.
27 versus 82.80.5.meraGPT Decider 1 meraGPT announced Decider 1 on September 22, 2026.It is a hosted, closed-weight, text-only model.It allows up to 10 Choice labels and a 4,096-token context.The model page lists $0.04 per 1M input tokens.It exposes /v1/systemone and works with the TypeSafe SDK.
On meraGPT’s typed-decisions run, it scored 0.768 versus Jev’s 0.727.An independent review notes its ECE of 0.180 trails Jev’s 0.144.6.Respan Span-01 Span-01 is a behavior-scoring model, not a general decision model.It returns present, absent or notobservable probabilities.It costs $0.
02 per 1M input tokens.Span-01 Lite is free.Respan claims 0.843 overall F1 on its Behavior Benchmark.Its own blog table shows a lower 0.806, with GPT-6 Sol ahead at 0.885.Size, base model and license are not disclosed.7.
Standard Thinking Standard One Standard One ships as 3B and 8B models, both open and hosted.The 8B is a LoRA on Ministral-3-8B under Apache-2.0.It reads text and images, with a 32,768-token context.It serves /v1/systemone.Hosted text pricing is $0.018 (3B) and $0.035 (8B) per 1M tokens.
On JevBench public hard, the 8B scored 54.95% versus Jev’s 72.07%.It led on realistic transfer, 90.83% versus 86.67%.Both runs are publisher-measured.Open-weight Jev alternatives you can self-host 8.Together AI Tev1-4B-experimental Together trained Tev1 as a LoRA SFT of Qwen3.5-4B.
It cost about $17 and 25 minutes on 37,840 examples.It picks from 2 to 24 options.The model card says the weights license is still being finalized.It answers through chat completions, not a native Jev endpoint.Respan measured 0.901 accuracy and 0.031 ECE, the best ECE in its table.
Together warns its logprobs are not calibrated confidence.9.Fastino GLiNER2.5-Decide Fastino’s GLiNER2.5-Decide is a 340M DeBERTa-v3-large encoder under Apache 2.0.It handles decisions, span extraction and relations in one pass.A 1B and a 287M multilingual variant also exist.
On Fastino’s 17-domain set, the model card lists 60.2% exact match.p50 latency is 167.3 ms on a 48-vCPU Xeon.It is not Jev API-compatible.10.Strands Labs Strands Decider 2B AWS’s Strands Labs introduced Strands Decider on October 1, 2026.It is about 1.9B: Qwen3.5-2B-Base plus a pointer head.
The repo publishes 246,678 training rows and the full recipe.One RTX 3090 trains it in about 11 hours.It serves a local /v1/systemone endpoint under Apache-2.0.The v1 checkpoint scored 180 of 231 (77.9%) on JevBench public.Median latency is 115 ms on an RTX 3090.11.Kev (Jared Palmer) Kev comes in 0.
8B, 4B, 9B and 27B sizes on Qwen3.5 and Qwen3.8.Kev-4B is a LoRA plus pointer head under Apache-2.0.It supports 255 options.The TypeSafe SDK works against it unchanged.Kev-4B fits a 24 GB GPU or a 32 GB Apple Silicon Mac.Respan measured 0.833 accuracy and a 0.060 flip rate.
Self-reported MMLU-Pro is 0.565 versus Jev’s 0.840.Option order can change its answers.12.Mapika Decider family Mapika’s Decider ships 0.8B to 35B-A3B variants on Qwen3.5 and Gemma 4.A decider-2b-vision model adds images.Everything is Apache-2.0, with 32k context and /v1/systemone.
On the JevBench public hard tier, decider-12b v1 tied Jev at 0.730.The 4B GGUF Q4K_M is 2.7 GB and runs on CPU.The 4B’s ECE is 0.184 on hard items, so expect overconfidence.13.Bespoke Labs Bespoke-Nimble-9B Bespoke-Nimble-9B is a 165 MiB LoRA on Qwen3.5-9B, Apache-2.0 on Hugging Face.
The current card lists 255 choices and 8,192 tokens.It reads English text only.On Bespoke’s 324-example held-out set, Nimble scored 90.12% versus Jev’s 93.21%.Latency is about 106 ms on an H100.It trained on only 2,676 examples, and its probabilities are not calibrated.14.
Convai Innovations Laya Laya is a ModernBERT-based family with 421M English and 322M multilingual checkpoints.It is Apache-2.0 and text only.The laya-serve server exposes a Jev-shaped /v1/systemone endpoint.The typed-decisions checkpoint scored 0.766 versus Jev’s 0.727.On Banking77, it scored 0.
425 versus 0.870.Respan measured 0.542 accuracy and a 0.396 injection success rate.p50 latency is 32.8 ms on a T4.15.Von (wfzyx) Von is a 395M ModernBERT-large encoder with an option-marker head.It scores all options in one bidirectional pass.It is Apache-2.0 with an 8,192-
Related
相關文章
Google 開源 Harness²：遞迴代理程式框架，無需新模型也能提升效能
Google Cloud AI Research 開源了 Harness²，這是一個遞迴代理程式框架，透過改良凍結模型周圍的程式碼與檔案來增強效能。對開發者而言，重點很簡單：你的代理程式無需更強大的模型也能進步。重點摘要：規模為框架而非模型，訓練參數為0，包含8個可編輯的框架元件，版本0.1.0。執行環境為Linux、Python 3.12或3.13搭配uv，模型呼叫則透過Vertex AI。GPU需求未公開。效能方面：一次改良步驟即可在3個基準測試中，提升所有4個模型與框架組合的表現。最佳成果：GLM-5搭配OpenCode在WorkBuddy-Bench上從64.2提升至75.7（+11.5點）；最差表現：GLM-5搭配Codex在JobBench上僅提升1.2點（23.6至24.8）。什麼是Harness²？它是Google研究推出的即插即用框架。
如何利用NVIDIA結構化資料模型進行零訓練表格預測
In this tutorial, we implement Structured Data Models, NVIDIA’s new library that puts several pretrained foundation models for tables behind one PyTorch interface, and use it to make predictions without training.
認識 Kyutai REKI：2.4B VLA 模型為機器人操控帶來高效記憶
Kyutai 與 Inria 及 École des Ponts 合作，推出了 REKI——一個用於機器人操控的 2.4B 參數視覺-語言-動作模型（VLA）。該模型僅使用公開數據進行預訓練，並配備了輕量級的近期記憶功能。在三個分佈外基準測試中，僅有規模約為其兩倍的模型能超越它。所有模型權重已於今日在 Hugging Face 上公開。核心資訊：規模：2.4B 參數（2.2B Qwen3.5-2B VLM + 0.2B 流匹配動作專家）。記憶涵蓋最近 3 幀，間隔 1 秒。運行環境：在 1 張 NVIDIA H100 上進行基準測試。從兩個 256×256 攝像頭輸入，編譯後每個動作區塊耗時 21.8 毫秒。權重以 BF16 格式提供，暫無量化版本。性能：在 LIBERO-Plus 和 RoboTwin Clean2Rand 基準測試中，為公開數據預訓練的 VLA 模型中最優；在 LIBERO-Plus 上達到 90.4% 的最高分。
DeepSeek 發布 V4.1-Flash-1017，DeepSWE 得分 77.0，並開放搜尋 API
DeepSeek 剛發布 DeepSeek-V4.1-Flash-1017，這是其 V4.1-Flash 模型的新檢查點，現已在 DeepSeek 網站、應用程式及 API 上線。DeepSeek 宣稱該版本在性能、任務完成速度與輸出簡潔度上均優於 9 月發布的 V4.1-Flash，並稱之為性能與成本的帕累託改進。同時，DeepSeek 也向所有開發者開放了搜尋與擷取 API。對代理建構者而言，這意味著更便宜的編碼模型加上單一 API 金鑰即可使用的第一方網路層。重點摘要：骨架參數 552B，另加 196B 的 Engram 參數；預填充階段每 token 啟用 8B 參數，解碼階段啟用 16B 參數；上下文長度達 1M tokens，最大輸出 384K tokens。運行於 DeepSeek API 的 deepseek-flash 模型。9 月版的 V4.1-Flash 權重（763B 參數，FP8）採 MIT 授權。
長鑫科技市值兩月內跌去1萬億；阿里加碼AI硬件品牌，天貓、速賣通聯動；國行版蘋果AI有望12月上線，百度、阿里參與合作丨雷峰早報
要聞提示1.長鑫科技市值兩月內跌去1萬億2.年內已至少18次故障！DeepSeek發放宕機補償，有人拿到3000元3.阿里加碼AI硬件品牌，天貓、速賣通聯動4.國行版蘋果AI有望12月上線，百度、阿里參與合作5.可靈AI新一代模型Kling 4.
Meta 發布 Muse Spark Open：816B 參數開源 MoE 模型，支援百萬級上下文
Meta 超級智慧實驗室正式釋出 Muse Spark Open，這是其 Muse Spark 模型的開源權重版本。Muse Spark Open 為一款 816B 參數的混合專家（MoE）模型，每個 token 啟動 42B 參數。它支援文字、圖片、影片與音訊輸入，輸出文字，並具備 100 萬 token 的上下文視窗。Meta 團隊宣稱，這是迄今為止美國效能最強的開源權重模型。對開發者而言，真正的重點在於部署：NVFP4 版本可在 2 臺配備 Apple M5 Ultra 的 Mac Studio 上運行。重點摘要：參數量為 816B 總參數／42B 啟動參數（256 個專家，每個 token 啟動 8 個）；100 萬 token 上下文。運行需求：透過 MLX 框架，2 臺 Mac Studio（M5 Ultra，各 512 GB 記憶體）；4 張 Blackwell Ultra GPU 或 8 張 H200（使用 NVFP4）；或透過 Unsloth 2-bit GGUF 量化，約需 300 GB 配置。效能表現：在代理式工具使用上表現強勁；在大型開源模型中位居中段班。