In collaboration with @nvidia we're releasing two new open weight models: Fastino-Nemotron-3.5-Lightning-Finance and Fastino-Nemotron-3.5-Lightning-Healthcare. Working closely with the Nemotron team, we developed both models on Nemotron 3.5 Lightning using the Fastino Fine-Tuning Agent, which autonomously ran the entire post-training pipeline. The base model was highly responsive, with the agent reaching substantial gains across both domains with compact datasets alone. - Fastino-Nemotron-3.5-Lightning-Finance excels at numerical reasoning, research, and summarization over financial concepts and documents. When evaluated, the fine-tuned model increased FinQA execution accuracy by 43.37 points (15.86% to 59.23%), and increased 7.81 points (from 49.65% to 57.46%) on BizFinBench. - Fastino-Nemotron-3.5-Lightning-Healthcare specializes in clinical conversation quality, summarizing notes and discharge documentation, and extracting medical concepts from unstructured text. In post-training, the Fastino Fine-Tuning Agent improved performance by a margin of 4.28 points on HealthAdminBench (from 25.67% to 29.95%) and 5.09 points on MedCalc-Bench (from 49.09% to 54.18%). It also increased flag accuracy on MEDEC by 11.32 points, beating Opus 4.6 and beating Muse Spark by 0.98 points. Both models are available on @huggingface under the Apache 2.0 license: - huggingface.co/fastino/Fasti… - huggingface.co/fastino/Fasti… Today we are also releasing the Fastino Fine-Tuning Agent in private preview, which you can sign up for access to here: fastino.ai/nvidia-collaborat…

A decision model that’s 90% confident isn’t necessarily right 90% of the time. When a model is uncalibrated, the confidence score reflects how strongly it prefers a label, but not how often that prediction is correct. Reinforcement Learning for Calibrated Decisions (RLCD) is a post-training technique that fixes this by rewarding models for making correct decisions and for expressing uncertainty accurately. The result is calibrated probabilities, so teams know when to automate a decision and when to escalate uncertain responses to a human reviewer. In our latest blog, we break down how RLCD works, why calibration matters for decision-making, and when to apply it to your own models. Read the full blog: fastino.ai/blog/rlcd-trainin…

LLMs are powerful, but they’re designed primarily for generating text rather than decision-making. Decision models are fast models that evaluate a given situation and return a label, choice, or confidence score used to take quick, decisive actions. With so many decision models available, it can be hard to know which one is right for your use case. We’ve broken the selection process down into 4 steps: 1. Define your possible outcomes, including options for uncertain cases. 2. Shortlist models based on task requirements and benchmark performance. 3. Compare accuracy, speed, and cost using the same test cases. 4. Set confidence thresholds to determine when human review is needed. In our recent blog, we break down popular use cases, some of the top decision models, and when you should train them. Read the full guide: fastino.ai/blog/decision-mod…

We changed how we train GLiDE, and it put us at #1 on the @huggingface Decision Index. Last week, we released the weights for GLiDE v1.2, the latest version of our System One decision model, and it scored 71.5 on the Hugging Face Decision Index 0.4.1. Here are some of the key changes that improved the performance: - Distilling reasoning: Instead of generating thinking tokens, we trained the model to learn from reasoning models, enabling decisions in a single forward pass. - Improving training data: Rather than adding more data, we restructured the training examples to better match the format of actual evaluation tasks. - Improving confidence calibration: By training the model to recognize when it doesn't have the answer, it reduced the amount of confident mistakes. In our latest blog, we go over what the entire training process looked like, some of the lessons we learned along the way, and the full training recipe we used. Read the full blog: fastino.ai/blog/training-gli… Download the model: huggingface.co/fastino/GLiDE

GLiDE 🏔️ on the accuracy/latency/cost pareto

just updated the Decision Index to 0.4 🧭 what's new? 🛡️ curbing benchmaxxing: score now exclusively made from private & in-house benchmarks 🔒 stronger private benchmarks on new domains 🏆 224 models evaluated, new top models in all sizes come play 👉huggingface.co/spaces/multim…

The best decision model in the world. GLiDE v1.2 is here (and it’s multimodal). #1 on the @huggingface Decision Index, beating the next best model by 1.7 points. The benchmarks are genuinely impressive: - 71.5 full score - 82 ms latency on NVIDIA RTX PRO 6000 - 1.9% calibration error on held-out decisions (one of the lowest on the index) This model is smart, fast, well-calibrated, and open weight. Everything a good decision model should be. We’ve also dropped our pricing to $0.04 / million input tokens. Try it today. Use promo code GLIDEIS1 for $10 in API credits. API: agent.fastino.ai/ Weights: huggingface.co/fastino/GLiDE

Our newest decision model is the world's best decision model. Introducing GLiDE v1.2, the #1 model on the @huggingface Decision Index. With an overall score of 71.5, GLiDE v1.2 also has one of the lowest calibration errors on the entire Decision Index, meaning its confidence more closely reflects how often its answers are correct. Beyond improved benchmark performance, it also adds support for multimodal inputs, enabling decision-making using text, JSON, and images. We are excited to release the model weights on Hugging Face. For hosted inference, you can also run GLiDE v1.2 through the Fastino API at $0.04 / million input tokens. Use promo code GLIDEIS1 for $10 in API credits. Model weights: huggingface.co/fastino/GLiDE API: agent.fastino.ai Blog: fastino.ai/blog/training-gli…

Open source models thrive when they’re built with feedback from the community. Sharing model weights and code allows developers to build on existing work, contribute improvements, and help the models evolve over time. In our latest video, founding researcher, @urchadeDS, shares how community feedback has helped shape GLiNER into what it is today. This week, we’re celebrating Open Source AI Week with the rest of the community in San Francisco. Open source has always been an important part of our work at Fastino, and it continues to play a major role in how and why we build. Watch the full conversation between @george_onx and Urchade: piped.video/6t6pZXsQXas

We're proud to be a post-training partner with @NVIDIAAI, advancing the future of open weight models.

You can now post-train Qwen3.6-35B-A3B with the Fastino Training API. Here are 3 ideas for what you can build: Decision model → FT on context-plus-options pairs labeled with the correct choice. RL rewards picking the right option with well-calibrated confidence in a single short answer. Text-to-SQL analyst → SFT on question-to-query pairs for a specific database schema. RL runs each query and rewards an exact match with the correct result set. Fast reasoner that thinks less → SFT on concise reasoning traces for math or code. RL rewards correct answers minus a penalty for thinking tokens, so it gets cheaper while maintaining accuracy. Training API: agent.fastino.ai

The Fastino Training API now supports Qwen3.6-35B-A3B for training and inference. It’s strong at coding and reasoning, with tool calling to plan and execute multi-step agentic workflows, making it a great foundation for production AI that needs to make decisions and take action. The model is also multimodal, supporting text, image, and video inputs. Start training: agent.fastino.ai

Our research team did a study and found that optimizing your harness can drastically improve model performance to a shocking extent. We ran frozen @NVIDIAAI Nemotron 3.5 Super VL weights through 9 off-the-shelf harnesses on the same 100 SWE-bench Verified tasks. The results ranged from 30 to 76 tasks solved, a huge spread. Each run used the same endpoint, 100 tasks, one attempt per task, and a 4,500-second limit, so the 46-task spread came from the harness. We found similar results with other models: - Claude Opus 4.5 went from 33.5% to 77.5% on BFCL V4 just by adding native tool schemas. - A new execution pipeline moved GPT-4o from 23.2% to 38.8% on SWE-bench Verified (but used much more inference). - AHE gained from tools + middleware + long-term memory on Terminal-Bench 2. Its system prompt alone hurt performance by 2.3 points. If you’re evaluating an agent, it’s worth benchmarking the model and harness together. The weights alone don’t tell you what the system can do. Read more: fastino.ai/blog/agent-harnes…

A frontier model can’t fix a poorly built harness. People assume swapping the underlying model for a smarter one will fix performance, but model intelligence is only one part of the equation. The harness surrounding the model can have just as much impact on whether that intelligence translates into reliable actions. It shapes how the model receives context, takes actions, maintains state, and verifies its work. After lots of experimentation with agent harnesses, our research team discovered a few key ways to improve their performance: 1. Give the model fewer, higher-level actions to choose from, and let the harness handle routine tasks. 2. Structure inputs and outputs using familiar formats and conventions the model has already seen in pretraining, instead of teaching it new ones. 3. Build rules and checks directly into the interface, such as rate limits and permissions, rather than relying on prompts to prevent mistakes. 4. Store important task information in durable storage rather than letting it accumulate in the context window, to keep the context focused on what’s most relevant. Before moving to a smarter model, check the surrounding software. A better harness could help you get more out of your model. Read Julia’s full guide to harness engineering: fastino.ai/blog/agent-harnes…

Big shoutout to all of our partners for making Nemotron 3.5 Super VL their own through post-training 🙌 Here’s what they’ve built 🧵

6 lessons from fine-tuning @NVIDIAAI Nemotron 3.5 Super VL across finance, healthcare and math: 1. Teach the interface; Super already has the knowledge -Largest gains from learning exact programs, spans + output fields -Small amounts of format-specific data made a huge difference -Focus on the artifact, not explanatory prose 2. A few hundred correctly labeled examples go a long way -500 examples produced a 42-point accuracy gain -Only 15 minutes of training -Correctness + consistent formatting mattered more than volume 3. Mix skills to protect neighboring tasks -Narrow tuning pulled related tasks toward the wrong output format -More single-skill data made the interference worse -A balanced mixture preserved multiple behaviors 4. Recipes transferred to Nemotron Super -Data + hyperparameters developed on Nemotron 3.5 Lightning reused unchanged -Transfer held across every task measured -Recipe search on the smaller model, then one run on Super 5. Fine-tuning teaches length, not just content -One task went from 1,597 → 113 output tokens -Accuracy still increased by 8.4 points -Math RL produced 25% shorter answers + ~6 points higher accuracy 6. Use 1x to 2x the full-adapter learning rate, and give the model room to think -2x LR produced larger healthcare gains on 7 of 8 evaluations -Higher token limits prevented strong answers from being cut off -At 1,024 tokens, 39% of base math responses were truncated For more on benchmarks and training recipes see our latest blog. Blog: fastino.ai/blog/nvidia-nemot…

We just released two fine-tuned @NVIDIAAI Nemotron 3.5 Super VL models that perform extremely well on complex agentic tasks in healthcare and finance domains. With 4.6x fewer active parameters, Super VL already outperforms Nemotron Ultra on long-horizon agentic tasks, making it a remarkably strong base for our models. Across finance and healthcare, every held-out benchmark we measured improved. The finance model: FinQA: 35.2% → 77.9% • TAT-QA: 68.5 → 87.8 F1 • SEC-Num: 79.2% → 93.9% • FinEntity: 84.7 → 92.4 • ConvFinQA: 22.3% → 76.3% (transfer) The healthcare model: BC5CDR: 63.8 → 84.8 F1 • BioRED: 46.1 → 69.0 (transfer) • MedMentions: 40.4 → 50.4 • PubMedQA: 70.2% → 78.6% • MedCalc-Bench Verified: 65.9% → 72.5% The transfer results are particularly encouraging. The finance model learned capabilities that carried into conversational financial reasoning and other unseen finance tasks. The healthcare model improved on BioRED despite training on a different biomedical corpus. These results use NVIDIA’s early-access checkpoint and text-only inputs, but they show how much domain-specific capability Nemotron 3.5 Super VL can gain through relatively small adaptations. And you can use the same API we used to fine-tune Nemotron 3.5 Super VL with per-token pricing up to 89% lower than Tinker’s. Start here: agent.fastino.ai/

Our team was given early access to @NVIDIAAI's new Nemotron 3.5 Super VL model to fine-tune two models specialized in finance and healthcare using the Fastino Training API. Both models demonstrated improved performance across all 15 benchmarks we evaluated, 5 of which included tasks the models had not been trained on. Fine-tuning a 76M-parameter adapter through the Fastino Training API improved the finance model’s performance on FinQA from 35% to 77% and cost just $2 to train. In our latest technical blog, we break down the results of our training, six lessons we learned along the way, and practical tips and training recipes for anyone looking to fine-tune the base model. Nemotron 3.5 Super VL is now available for training and inference on the Fastino Training API, with a limited-time discount for training. API: agent.fastino.ai/ Blog & Recipes: fastino.ai/blog/nvidia-nemot… Model weights: huggingface.co/collections/f…

Everyone is talking about specialized models today, so perfect timing to show this! I gave the Fastino Agent one goal: teach GLiNER Decide to play Mario. It built data, fine-tuned via our API, analyzed failures, and ran each new experiment autonomously. The before/after is wild

@NVIDIAAI Nemotron 3.5 Super VL launched today. You can already post-train it through the new @fastinoAI Training API. You control the data, rewards, evals, and training loop. API handles sampling, gradients, weight updates, and checkpoints. The API offers: -SFT + RL -PPO, CISPO, DRO + importance sampling -Tinker SDK + cookbook compatible -Per-token pricing, no idle GPU charges -Lower rates on comparable models + token types -Nemotron 3.5 Lightning support We’re also releasing finance and healthcare fine-tunes. Better than the base model across all 15 held-out benchmarks we measured, trained in 2 to 6 hours. API: agent.fastino.ai/