Introducing the Decision Index 0.1 ⚖️ a rigorous leaderboard comparing jev with 30+ open weights decision models 35+ benchmarks. asking 130K questions to each model testing knowledge 🧠, automation ⚙️, understanding 🤔and even creativity 🎨

Happy to officially bring you the new SOTA `tokenization` library. We focused on all languages, multi-thread scaling, minimal package size and memory usage. We're thankful for the players of this ecosystem that have pushed us to give the best we could!

2 new OCR models landed on @huggingface today - tencent/WeVisDoc (2B and 4B), Apache 2.0 - jinaai/jina-ocr-v1, non-commercial license Which one should you pick? Based on the OmniDocBench v1.6 benchmark on Papers with Code, WeVisDoc-4B leads with an overall score of 95.38, but the current SOTA is NaviDC-OCR

Introducing the Papers with Code MCP server! To celebrate its release, I've asked Claude Code to research the architecture of Jev by @typesafeai using the `search_papers` tool. Here's what it came up with: "If Jev is architecturally anything public, it's a large schema-conditioned bidirectional encoder with parallel label-query heads (GLiNER/ML-Decoder shape) trained with a proper-scoring-rule RL objective (RLCR shape), scaled far past the ~150M-param range those papers operate in — "neither small nor an LLM" fits that. The 40–200× speedup is consistent with removing autoregression, not with any exotic mechanism."

Dear friends of @cursor_ai, @mntruell I'm very surprised at how bad Cursor has become... It became my favorite AI tool after @code, and I used it non-stop. The UX was so much better than any other tool out there. However, after getting Codex for open-source, I tried Codex Desktop and basically didn't look back, also thanks to @thsottiaux. However, given that my Codex credits are depleted from time to time, I return to Cursor, and each time I'm surprised at how bad it is compared to Codex Desktop. The product seems heavily vibe-coded, with various basic things not working, like: - chats not syncing between the IDE and the Agents Window app (?) - clicking "Open Workspace" and it just doesn't open it - very weird UI issues like the one below with giant whitespace between tool calls. I tweeted about it a month ago. Someone on your team then reached out to me to fix these issues, but we're a month later, and things are still broken (I'm on 3.20.21). I don't know what happened. Maybe the focus is on @bot now? I'm happy to give all my feedback to improve the product. I'm just sad to see it going this way.

Nearly half a year of silence. We spent it studying one problem: how far RL can scale. MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks. Streaming the run:

For folks wondering what YOCO means, you can simply ask it on Papers with Code :) It powers the new DeepSeek-V4.1-Flash architecture. It's not an encoder-decoder, but rather a decoder-decoder architecture. It's all about reducing GPU memory and prefill latency. Find the chat here:

For folks wondering what Sliding Window Attention is, there's a method for it on Papers with Code Sliding Window Attention (SWA): A local attention pattern that restricts each token to attending only within a fixed-size neighborhood instead of the full sequence. This reduces attention and KV-cache memory for long-context models, while periodic global-attention layers can preserve broader context. Find it here:

People have reported issues on to me Apologies! Working on a more scalable architecture Should I just open-source the code base so we can collectively improve it?