Seeing Jev all over my feed got me curious, so I tried the same idea with LFM2.5-VL on a driving video. Just a small demo to test visual decision readouts from first token probabilities. Not perfect and the latency is definitely there 😅

TensorFold 1.2.5 is out. 🚀 This release brings: • TensorFold Web, a local browser app for chat, models, learning, benchmarks and more. • LFM2.5-24B-A2B on CUDA. • Nemotron Parse and LightOnOCR-3 on HIP. • Clef's joint schema head on CUDA. Qwen3.8-27B MLX 8-bit on M5 Ultra also gets faster. In the measured test, cold prompts were 18% faster and decode 7.6% faster, with the same tokens and prompt state. There's lower memory use for Qwen CUDA streams, better conversation-cache handling and a fix for PDF rendering crashes on macOS too. A massive thank you to everyone contributing code, testing builds and reporting problems. These releases take a lot of work, and I really appreciate the time people put into them. @Bizuayeu @devopsfortunato @unigilby @teleogen @dubsthedev @kalin_swiss @Q4_K_M And on GitHub: AbrahamCete, gordonhatusupy, jschmied, mrv777 and yguanrong. Special credit to Stephane on GitHub for TensorFold Studio, which TensorFold Web builds on. @volatilemarkts @plotarmordev @MiaAI_lab @petruspennanen

The same model, with the same weights, scores 62% in one agent harness and 33% in another. @adithya_s_k and the @huggingface team just released the ultimate guide to multi-harness RL, and it's one of the most practical RL write-ups this year, and everything open! The trick is simple. Don't touch the harness. Point it at a proxy instead of the model. The proxy speaks all four API formats coding agents use (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Gemini). It records the exact token ids and logprobs vLLM sampled, and you train on that. You don't change a single line of Claude Code, Codex or OpenCode. Results: 🔹 Trained across 4 harnesses at once, LFM2.5-2.6B by @liquidai went from 42% to 54% 🔹 31% fewer tool calls, thanks to a small bonus for solving tasks in fewer steps 🔹 Training in OpenCode alone took OpenCode from 34% to 58%, but the multi-harness model improved everywhere They also tried the shortcut everyone reaches for: fine-tune on 3,189 successful rollouts from Qwen3.8-27B. Imitation plateaued at 47.5%, below both RL runs. Copying a bigger model doesn't get you there. Practice does. The best part is that everything is open: the capture proxy in OpenEnv, the trainer in TRL, the tasks, the SFT data, the training code and all seven trained models. Agents will run in dozens of harnesses. Now open models can be trained for each of them, by anyone. Read it here 👇

been locked in trying to set up a crude continual learning environment that I can develop into a personal continuous learning bench. nothing works yet, but here's the plan: - take some daily tasks that agents can do but lend themself to customisation, like daily planning. - build an rl environment around the daily tasks - add synthetic preferences to the environment. i.e. how i like my day planned - get a judge model to judge traces based on preferences - get agents to do the tasks and record the traces as buckets - benchmark the base model on the task (LiquidAI/LFM2.5-1.2B-Instruct) - SFT the model on the traces - SDPO a model on the traces with a judge model hint So far, Evals work, SFT works, but SDPO doesn't. Pretty sure that the preferences are too arbitrary, so considering manually adding more preferences to setup a better few shot judge.

The new LFM 2.5 2.6B model by @liquidai is now available on iPhone and iPad. A fast model designed for edge devices with comparable or better scores compared to models up to nearly 4× its size. Update your app now.

The new LFM 2.5 VL 450M model by @liquidai is now available in the app. A lightning-fast and lightweight model perfect for image understanding in shortcuts. Update your app now.