DeepOpen
Open multilingual non-autoregressive System 1 engine: a script-aware Router picks English, multilingual, or typed-decisions checkpoints, then one forward pass returns Choice, Score, and Noul.
Overview
DeepOpen (github.com/deepopen-com/deepopen, Apache-2.0, pip install deepopen) is an open multilingual non-autoregressive System 1 decision engine built on Laya. 1 forward pass answers choice, score, and noul over text, email, ticket, or JSON state. The Router detects script and language in pure Python, under 0.5 ms per the README, and sends the request to one of 3 Hugging Face checkpoints: convaiinnovations/deepopen (ModernBERT-large, 421M, 512 context, English), convaiinnovations/deepopen-multilingual (mmBERT-base, 322M, 1024 context, 100+ languages, README says about 2x the English checkpoint on speed), and convaiinnovations/deepopen-typed-decisions (ModernBERT-large, 421M, 1024 context, fine-tuned on typed workflows). You can also call deepopen.load on the root repo with subfolder multilingual or typed-decisions. README speed tables on a Tesla T4 put one multilingual question at 32.8 ms and a 10-question batch at 72.3 ms (7.2 ms per question). Preload keeps checkpoints resident. The default max_loaded of 1 rebuilds a model on every language switch, measured at a 7.4 s median on CPU and 10.3 s on T4. The GitHub repo lists www.deepopen.com as its homepage. Public counts were 1,633 stars and 217 forks when this listing was drafted.
Problem: English-only open encoders and a single hosted System One endpoint both miss the case where the text is not Latin script, the checkpoint has to change before the forward pass, and you still want Choice, Score, and Noul probabilities you can threshold.
Built for: Python teams who want an Apache-2.0 local decision stack with an automatic English, multilingual, or typed-decisions router, and who will read the README limits on high-cardinality labels and zero-shot typed workflows.
First indexed on Jev Directory: 2026-10-01
Creator and team
- Name
- Deep Open
- Organization
- Deep Open
- Handle
- @deepopen-com
“The built-in Router is the recommended entry point: it evaluates any state in any language, automatically detects scripts and languages in sub-milliseconds, and dispatches to the optimal checkpoint in a single forward pass.”
How Jev is used
- Role in the product flow
- Script-aware Router picks one of 3 non-autoregressive checkpoints, then 1 forward pass answers Choice, Score, and Noul
- Primitives
- ChoiceScoreNoul
- State in
- Any state object (email fields, ticket text, or JSON) plus a questions map of choice criteria, score rubrics, or noul instructions. router.route inspects script and language without a forward pass.
- Decision out
- answers with choice label and confidence, score expectation, or noul probability, plus routing metadata (model, repo, reason). Preset packs cover model routing, prompt guardrails, moderation, and ticket triage.
- pip install deepopen, then Router(preload=True) or deepopen.load on convaiinnovations/deepopen
- Pass state and a questions map of choice, score, and noul items, or a preset such as triage_questions()
- Router detects script in pure Python and dispatches to english, multilingual, or an explicit typed-decisions override
- 1 forward pass returns answers and routing.reason. README example gates automated action at confidence 0.85 and escalates the rest
- For a dedicated pipeline, load one subfolder and skip the router. Raise head_max_len when a choice has dozens of options
DeepOpen is the multilingual router plus 3 checkpoints, built on Laya. The product is that router and those weights, not a second copy of an existing open model page. wfzyx-von is the English encoder aimed at order-invariant option scores and local speed. receptron-laya is the Node ONNX runtime for upstream Laya weights, and these 3 DeepOpen checkpoints are a different release line on convaiinnovations. nokia-applied-research-anyjev fits a Decider on hub LLMs you already serve and prints BANKING77 ECE. tianyucodings-nanojev trains 0.6B parallel heads on public game tasks. theoleecj-semif documents another open System One server. ollaya-dev-ollaya pulls and serves laya, kev, von, and decider ONNX on port 11435. githubnext-localjev bridges chat JSON on a Mac. rizzo-ai-academy-rizzo-flow serves a Spark head through llama.cpp. Hosted TypeSafe Jev remains the closed API. README comparison on 2,000 typed decisions: deepopen-typed-decisions argmax 0.766 against published Jev 1.13.0 at 0.727, with soft accuracy still behind at 0.471 against 0.580, and raw ECE 0.213 against 0.144 before temperature fitting. A separate README row lists post-temperature ECE at 0.081 against a third-party Jev ECE of 0.246. The same README gives Jev Banking77, 0.870 on 72 labels against 0.425 on 77 labels at the default head budget, and says the base checkpoints sit near chance on the typed-decisions set until you fine-tune. The Fine-Tuning section describes a Kaggle 2xT4 RLCD loop, about 4 to 5 hours for 4 epochs over about 30k questions. The notebook in the tree is notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb. Reproduction trees live at banking77/ and clinc150/. Cite github.com/deepopen-com/deepopen README, BENCHMARKS.md, those trees, and the Hugging Face repos convaiinnovations/deepopen, deepopen-multilingual, and deepopen-typed-decisions.
Sourced performance claims
- Checkpoint table: deepopen is ModernBERT-large, 421M, 512 context. deepopen-multilingual is mmBERT-base, 322M, 1024 context. deepopen-typed-decisions is ModernBERT-large, 421M, 1024 context.Source: github.com/deepopen-com/deepopen README checkpoint table
- Tesla T4 speed table: multilingual 32.8 ms for 1 question, 40.1 ms for 5, 72.3 ms for 10 (7.2 ms per question), 337 ms for 50. English checkpoint: 39.5 ms, 84.5 ms, 158.6 ms, 771 ms on the same rows. Batched throughput 103 to 332 questions per second on one T4.Source: github.com/deepopen-com/deepopen README Speed (Tesla T4, measured)
- typed-decisions, 400 cases and 2,000 decisions: deepopen-typed-decisions accuracy 0.766, soft accuracy 0.471, Brier 0.062, ECE 0.213, score MAE 0.242. Published Jev 1.13.0 on the same table: 0.727, 0.580, 0.148, 0.144, 0.391. Base checkpoints 0.362 and 0.342 sit under the 0.461 majority-class baseline. README says the 0.766 figure is the fine-tune, and that Jev still leads soft accuracy and raw ECE before temperature fitting.Source: github.com/deepopen-com/deepopen README typed-decisions table and Honest limits
- Shared route table, 17,416 questions: Router keeps English MASSIVE intent at 0.783 and lifts 13 other languages from 0.306 to 0.451. XNLI English 0.860, 14 other languages 0.731. Languages usable above 3x random: 45 of 51. Khmer on the English checkpoint: 0.000 accuracy at 0.952 confidence. Banking77: Jev 0.870 on 72 labels, deepopen 0.425 on 77 labels at the default head budget.Source: github.com/deepopen-com/deepopen README Why Route and Where Jev leads
- Public GitHub repo deepopen-com/deepopen had 1,633 stars and 217 forks on 2026-10-01. License Apache-2.0. Last push 2026-09-28.Source: GitHub API 2026-10-01
Features and stack
Features
- Router with preload, max_loaded, attach, and unload
- 3 checkpoints, plus direct deepopen.load with subfolders
- choice, score, and noul in 1 forward pass, with routing.reason
- Preset packs: router_questions, guard_questions, moderation_questions, triage_questions
- Banking77 and CLINC150 reproduction directories
- BENCHMARKS.md plus temperature-calibration notes and a typed-decisions fine-tune notebook
Stack
- Python
- Apache-2.0
- ModernBERT-large and mmBERT-base
- Hugging Face convaiinnovations/deepopen
- PyPI deepopen
Pricing: Apache-2.0 weights. Inference cost is your GPU or CPU. README contrasts self-hosting with TypeSafe list price of $0.042 per 1M tokens for hosted Jev, and notes Jev figures in the comparison table are third-party published numbers.
Links
FAQ
- Is this the same product as receptron/laya or Von?
- No. receptron/laya is a Node ONNX client for upstream Laya weights. Von is an English non-autoregressive encoder focused on order-invariant scores. DeepOpen ships 3 checkpoints and a script-aware Router. The README says it is built on Laya.
- Does 0.766 mean DeepOpen beats hosted Jev with no fine-tune?
- README Honest limits: the base checkpoints score about 0.36 on typed-decisions, under a 0.461 majority-class baseline. The 0.766 accuracy is deepopen-typed-decisions, fine-tuned on that benchmark's training split. Soft accuracy on the same table is 0.471 against published Jev 0.580.
- Where does the README say hosted Jev still leads?
- High-cardinality choice. Banking77 is 0.870 for Jev on 72 labels and 0.425 for deepopen on 77 labels at the default head budget. README says options share head_max_len, so 77 labels get only a few tokens each unless you raise head_max_len or split the choice.
- Why preload the Router?
- README: a cold checkpoint build costs seconds. At max_loaded 1, alternating languages reloads a model every request, measured at 7.4 s median on CPU and 10.3 s on T4. Router(preload=True) keeps checkpoints resident. Language detection stays under 1 ms, and steady per-request latency is 32.8 ms on GPU or 193 to 464 ms on CPU.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Where to run Jev
Compare official TypeSafe, Vercel AI Gateway, OpenRouter, Cloudflare Workers AI, and classifier.dev with a fact table.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.