Boston police say they received a call about an individual walking around with what looked like a rifle or long gun, but officers have yet to find someone matching that description.

East Boston schools put into safe mode after report of armed person in area

Bitterness over the contentious legal action was evident from both sides this week.

Google to purge 'private' Incognito mode user records, but will keep snooping

Google has agreed to settle a lawsuit from consumer who alleged the company tracked user data even in private browsing <strong>mode</strong>.

Google agrees to delete billions of Incognito mode data records | Mashable

MODE

Raw measurements don’t tell the full story. 

Why Measuring Time is Not Enough: a Practical Roofline Model for ML Training

This tutorial shows you how to build a WhatsApp bot that accepts images, analyzes them using AI vision models, and responds with intelligent insights. 

How to Build an AI-Powered WhatsApp Bot That Analyzes Images Using Python and Vision Models

ThinkSafe shows why reasoning-optimized LLMs get less safe—and how self-generated safety traces plus refusal steering restore safety without killing reasoning.

ThinkSafe: Make Reasoning Models Safer Without Breaking Their Brain

<p><em>This is a simplified guide to an AI model called <a href="https://www.aimodels.fyi/models/huggingFace/glm-image-zai-org?utm_source=hackernoon&utm_medium=referral">GLM-Image</a> maintained by <a href="https://www.aimodels.fyi/creators/huggingFace/zai-org?utm_source=hackernoon&utm_medium=referral">zai-org</a>. If you like these kinds of analysis, join <a href="https://www.aimodels.fyi/?utm_source=hackernoon&utm_medium=referral">AIModels.fyi</a> or follow us on <a href="https://x.com/aimodelsfyi">Twitter</a>.</em></p>

<h2>Model overview</h2>
<p><code>GLM-Image</code> combines an autoregressive generator with a diffusion decoder to create images from text and perform sophisticated image editing tasks. The 9B autoregressive component generates a compact visual token sequence, which the 7B diffusion decoder then expands into high-resolution outputs. This hybrid architecture delivers competitive image quality while excelling in scenarios demanding precise text rendering and knowledge-intensive details. Unlike pure diffusion models that struggle with text accuracy, this approach handles complex information expression with greater fidelity. The model maintains the semantic understanding capabilities of <a href="https://aimodels.fyi/models/huggingFace/glm-4.1v-9b-thinking-zai-org?utm_source=hackernoon&utm_medium=referral">GLM-4.1V-9B-Thinking</a> while specializing in visual generation rather than multimodal reasoning.</p>
<h2>Model inputs and outputs</h2>
<p><code>GLM-Image</code> processes both textual prompts and image inputs to generate or modify visual content. The system accepts text descriptions of arbitrary length and complexity, along with optional reference images for conditional generation tasks. The model outputs high-resolution images ranging from 1024x1024 to 2048x1024 pixels, maintaining visual coherence and detail quality across diverse generation scenarios.</p>
<h3>Inputs</h3>
<ul>
<li><strong>Text prompts</strong> describing desired image content, including detailed specifications for layout, composition, and visual elements</li>
<li><strong>Reference images</strong> for image-to-image tasks such as editing and style transfer</li>
<li><strong>Multiple reference images</strong> for maintaining consistency across subjects in multi-image generation scenarios</li>
<li><strong>Guidance parameters</strong> controlling the strength of prompt adherence and generation diversity</li>
</ul>
<h3>Outputs</h3>
<ul>
<li><strong>Generated images</strong> in PNG format at specified dimensions</li>
<li><strong>High-fidelity visual content</strong> with accurate text rendering and fine-grained details</li>
<li><strong>Edited or transformed images</strong> preserving identity and compositional elements from source images</li>
</ul>
<h2>Capabilities</h2>
<p>The model generates images from text descriptions with particular strength in information-dense layouts. Recipe illustrations with ingredient lists, numbered steps, and visual hierarchies render with precise text and accurate proportions. Image-to-image capabilities include background replacement, style transfer between images, identity-preserving generation for people and objects, and multi-subject consistency maintenance. The decoupled reinforcement learning training using the GRPO algorithm enhances both semantic alignment and visual detail quality, with the autoregressive module improving aesthetic understanding and the diffusion decoder refining texture fidelity and text accuracy. Post-training feedback signals target low-frequency features like composition and high-frequency elements like texture, creating nuanced control over output characteristics.</p>
<h2>What can I use it for?</h2>
<p>Design professionals can generate detailed marketing materials with embedded text and complex layouts without manual typography adjustments. Educational content creators produce recipe cards, instructional guides, and infographics with embedded text that renders correctly on the first attempt. E-commerce businesses create product photography variations with consistent branding and identity. Visual artists explore style transfer to apply artistic treatments while preserving subject identity. Publishers generate magazine layouts and illustrated guides combining dense information with high visual quality. The model's accuracy with text-based content opens applications previously requiring manual post-processing or graphic design expertise. Compare this capability to <a href="https://aimodels.fyi/models/huggingFace/z-image-turbo-tongyi-mai?utm_source=hackernoon&utm_medium=referral">Z-Image-Turbo</a> for efficiency-focused workflows or <a href="https://aimodels.fyi/models/huggingFace/visualglm-6b-zai-org?utm_source=hackernoon&utm_medium=referral">visualglm-6b</a> for combined understanding and generation tasks.</p>
<h2>Things to try</h2>
<p>Test the model's text rendering accuracy with prompts containing typography-heavy content like posters, menus, or certificates. Generate illustrated recipes or instructional manuals with step-by-step visual guides and embedded text descriptions. Perform background replacement while preserving foreground subjects, then compare results when providing multiple reference images for consistency constraints. Apply artistic styles to photographs while maintaining face identity for personal portrait variations. Create magazine-style layouts with multiple content sections, experimenting with guidance scale values to balance creative interpretation against strict prompt adherence. The Glyph Encoder text module distinguishes this model from standard diffusion approaches, making typography-intensive generation a particularly revealing test case.</p>

<hr/><p><strong>Original post:</strong> <a href="https://www.aimodels.fyi/models/huggingFace/glm-image-zai-org?utm_source=hackernoon&utm_medium=referral">Read on AIModels.fyi</a></p>

Meet GLM-Image: A Hybrid Text-to-Image Model Built for Typography, Layouts, and Dense Info

Simulation results show how cooperative cyber insurance and smart monitoring reduce power grid risk under epidemic-style cyberattacks.

Cyber Insurance Models Show Reliability Gains for Power Grids Under Epidemic-Style Attacks

A new epidemic-based cyber-physical model links power grid cyberattacks to load loss, reliability, and insurance pricing strategies. 

Modeling Cyberattacks on Power Grids to Price Risk in Mutual Insurance Systems

A new cyber-insurance model uses smart monitoring, epidemic risk modeling, and Shapley values to fairly price cyber risk in power grids.

A New Cyber-Insurance Model Uses Game Theory to Protect Smart Power Grids

Discover parameter estimation for wrapped Cauchy and von Mises distributions in trajectory learning.

Swarms on Manifolds for Deep Learning: Training Kuramoto Models and Trajectory Learning

Master statistical ML on Grassmannians and orthogonal groups. Learn how matrix Bingham and matrix Langevin distributions derive from vMF spherical models.

Probabilistic ML on Grassmannians and Orthogonal Groups: Langevin and Bingham Matrix Models

Discover how statistical models over hyperbolic spaces enable inference, sampling, and density estimation in Geometric Deep Learning

Hyperbolic Space Statistical Models: Geometric Deep Learning & Inference

After 7 major replatformings over 11 years, I've cataloged 5 hidden failure modes that silently break database migrations weeks after everyone stops watching.

Why Most Database Migrations Fail Silently: 5 Hidden Failure Modes in Zero-Downtime Replatforming

<p><em>This is a simplified guide to an AI model called <a href="https://www.aimodels.fyi/models/huggingFace/iquest-coder-v1-40b-loop-instruct-iquestlab?utm_source=hackernoon&utm_medium=referral">IQuest-Coder-V1-40B-Loop-Instruct</a> maintained by <a href="https://www.aimodels.fyi/creators/huggingFace/IQuestLab?utm_source=hackernoon&utm_medium=referral">IQuestLab</a>. If you like these kinds of analysis, join <a href="https://www.aimodels.fyi/?utm_source=hackernoon&utm_medium=referral">AIModels.fyi</a> or follow us on <a href="https://x.com/aimodelsfyi">Twitter</a>.</em></p>

<h2>Model overview</h2>
<p><code>IQuest-Coder-V1-40B-Loop-Instruct</code> represents a specialized variant of the IQuest-Coder-V1 family, a collection of code language models built by <a href="https://aimodels.fyi/creators/huggingFace/IQuestLab?utm_source=hackernoon&utm_medium=referral">IQuestLab</a> for advanced software engineering tasks. This 40 billion parameter model uses a recurrent transformer design that processes information through two iterations, allowing it to balance model capacity with practical deployment requirements. The Loop variant achieves competitive performance on major coding benchmarks including SWE-Bench Verified (81.4%), BigCodeBench (49.9%), and LiveCodeBench v6 (81.1%), placing it alongside other leading code models like the standard <a href="https://aimodels.fyi/models/huggingFace/iquest-coder-v1-40b-instruct-iquestlab?utm_source=hackernoon&utm_medium=referral">40B Instruct variant</a>.</p>
<h2>Model inputs and outputs</h2>
<p>The model accepts natural language instructions and code-related prompts, processing them through its 128K token context window to generate coherent code solutions and explanations. It produces code snippets, complete implementations, debugging assistance, and technical explanations based on user queries.</p>
<h3>Inputs</h3>
<ul>
<li><strong>Natural language instructions</strong>: Requests for code generation, refactoring, or debugging</li>
<li><strong>Code snippets</strong>: Partial implementations to complete or improve</li>
<li><strong>Programming problems</strong>: Tasks requiring algorithmic solutions</li>
<li><strong>Chat messages</strong>: Multi-turn conversations formatted for instruction-following</li>
</ul>
<h3>Outputs</h3>
<ul>
<li><strong>Generated code</strong>: Complete, working implementations in multiple programming languages</li>
<li><strong>Code explanations</strong>: Detailed breakdowns of logic and functionality</li>
<li><strong>Debugging suggestions</strong>: Fixes for errors and optimization recommendations</li>
<li><strong>Technical documentation</strong>: Comments and explanations embedded in generated code</li>
</ul>
<h2>Capabilities</h2>
<p>The model handles complex coding scenarios through its code-flow training approach, which learns from repository evolution patterns and real-world development workflows rather than treating code as static text. It supports up to 128K tokens natively, enabling it to work with entire files, multiple related modules, or detailed project context without requiring additional scaling techniques. The Loop mechanism allows the model to refine its outputs iteratively, improving solution quality within a single generation pass.</p>
<p>Temperature settings of 0.6, TopP of 0.85, and TopK of 20 are recommended for optimal performance. The model demonstrates strong reasoning capabilities for competitive programming problems, tool usage in agentic software engineering scenarios, and general-purpose coding assistance tasks.</p>
<h2>What can I use it for?</h2>
<p>Organizations can deploy this model for autonomous code generation pipelines, reducing development time on routine implementation tasks. Software engineering teams can use it as a pair programming assistant that understands repository-scale context, helping developers navigate unfamiliar codebases or generate boilerplate code. The model serves well in educational settings where students need immediate feedback on coding exercises, or in enterprise environments building internal coding assistants. Companies building developer tools can integrate it via the vLLM framework to create OpenAI-compatible APIs, monetizing through usage-based pricing or subscription tiers.</p>
<h2>Things to try</h2>
<p>Experiment with providing the model large code repositories or entire project directories within the 128K context window to see how effectively it understands interconnected code relationships and generates implementations that maintain consistency across multiple files. Test its performance on complex refactoring tasks where understanding the broader codebase matters more than isolated code generation. Try using it iteratively within a single conversation to progressively improve and debug implementations, leveraging its instruction-following design for multi-step problem solving. Compare its outputs against the standard Instruct variant to understand the practical differences the Loop mechanism introduces in solution quality versus generation speed.</p>

<hr/><p><strong>Original post:</strong> <a href="https://www.aimodels.fyi/models/huggingFace/iquest-coder-v1-40b-loop-instruct-iquestlab?utm_source=hackernoon&utm_medium=referral">Read on AIModels.fyi</a></p>

This 40B “Loop” Coder Model Tries Twice So Your Code Doesn’t Have To

<p><em>This is a simplified guide to an AI model called <a href="https://www.aimodels.fyi/models/huggingFace/llama3.3-8b-instruct-thinking-claude-4.5-opus-high-reasoning-davidau?utm_source=hackernoon&utm_medium=referral">Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-Reasoning</a> maintained by <a href="https://www.aimodels.fyi/creators/huggingFace/DavidAU?utm_source=hackernoon&utm_medium=referral">DavidAU</a>. If you like these kinds of analysis, join <a href="https://www.aimodels.fyi/?utm_source=hackernoon&utm_medium=referral">AIModels.fyi</a> or follow us on <a href="https://x.com/aimodelsfyi">Twitter</a>.</em></p>

<h2>Model overview</h2>
<p><code>Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-Reasoning</code> combines an unreleased Llama 3.3 8B base model with extended 128k context capabilities and specialized thinking/reasoning training. The model was fine-tuned using Unsloth over three epochs with the Claude 4.5-Opus High Reasoning dataset, creating a hybrid instruct and thinking model. Created by <a href="https://aimodels.fyi/creators/huggingFace/DavidAU?utm_source=hackernoon&utm_medium=referral">DavidAU</a>, this model activates extended reasoning modes based on prompt content while maintaining instruction-following capabilities. The approach differs from similar models like <a href="https://aimodels.fyi/models/huggingFace/llama-3.2-8x4b-moe-v2-dark-champion-instruct-uncensored-abliterated-21b-gguf-davidau?utm_source=hackernoon&utm_medium=referral">Llama-3.2-8X4B-MOE-V2-Dark-Champion-Instruct-uncensored-abliterated-21B-GGUF</a> and <a href="https://aimodels.fyi/models/huggingFace/llama-3.2-8x3b-moe-dark-champion-instruct-uncensored-abliterated-18.4b-gguf-davidau?utm_source=hackernoon&utm_medium=referral">Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF</a> by focusing on reasoning activation rather than mixture-of-experts scaling.</p>
<h2>Model inputs and outputs</h2>
<p>The model accepts text prompts and generates responses with automatic thinking tag insertion when reasoning tasks are detected. Context windows of 8k or larger are recommended, though the model functions with 4k minimum. Specific trigger phrases such as "Think Deeply," "Explain," and "Tell me a horror story" activate extended thinking modes where the model performs multi-stage reasoning before generating output.</p>
<h3>Inputs</h3>
<ul>
<li><strong>Text prompts</strong> with or without explicit reasoning indicators</li>
<li><strong>Context</strong>: Supports up to 128k token context window</li>
<li><strong>Configuration parameters</strong>: Temperature (suggested 0.7), repetition penalty (1.05), top-p (0.95), min-p (0.05), top-k (40)</li>
</ul>
<h3>Outputs</h3>
<ul>
<li><strong>Thinking sequences</strong>: Tagged reasoning blocks that show internal problem-solving steps</li>
<li><strong>Final responses</strong>: Generated text following thinking phases or direct instruction responses</li>
<li><strong>Extended format</strong>: Supports creative writing, technical explanations, and narrative content</li>
</ul>
<h2>Capabilities</h2>
<p>The model demonstrates strong performance in technical reasoning tasks, generating detailed mathematical explanations and examples with proper derivations. When given prompts requesting deep analysis, it produces extended thinking sequences before formulating responses. Creative writing and story generation benefit from the reasoning training, producing narratives with coherent structure and thematic development. The 128k context enables handling of substantial documents and complex multi-part requests. Quantization options like Q4KS and IQ3_M maintain reasoning activation, though lower quantizations may degrade thinking performance.</p>
<h2>What can I use it for?</h2>
<p>Technical education and explanation generation work well for subjects like orbital mechanics, physics, and mathematics where step-by-step derivations enhance understanding. Creative fiction and storytelling benefit from the thinking phase, producing more carefully constructed narratives. Research and analysis tasks leverage the extended reasoning capability to break down complex problems. The model suits local deployment scenarios where users need reasoning without external API calls. Unlike models purely focused on creative outputs like <a href="https://aimodels.fyi/models/huggingFace/l3.2-rogue-creative-instruct-uncensored-abliterated-7b-gguf-davidau?utm_source=hackernoon&utm_medium=referral">L3.2-Rogue-Creative-Instruct-Uncensored-Abliterated-7B-GGUF</a>, this model balances reasoning depth with instruction adherence.</p>
<h2>Things to try</h2>
<p>Experiment with prompts that include phrases like "Think deeply" or "Explain in detail" to activate the reasoning mode and observe the thinking process for complex problems. Test the model's performance on multi-step mathematical problems where the reasoning output illuminates the solution path. Try creative writing prompts without thinking activation indicators to see how instruction mode alone handles narrative generation. Adjust the smoothing factor to 1.5 in supported interfaces like KoboldCpp or text-generation-webui for improved chat and roleplay consistency. Test different quantization levels starting with Q4KS to find the balance between reasoning quality and inference speed on your hardware.</p>

<hr/><p><strong>Original post:</strong> <a href="https://www.aimodels.fyi/models/huggingFace/llama3.3-8b-instruct-thinking-claude-4.5-opus-high-reasoning-davidau?utm_source=hackernoon&utm_medium=referral">Read on AIModels.fyi</a></p>

This Llama Model Knows When to Think: Inside Llama3.3-8B Instruct + “Reasoning Mode”

AI’s biggest bottleneck isn’t code, it’s infrastructure. Power, compute and governance now define who can actually scale AI.

AI Power Isn’t Just About “Better Models.” It’s About Who Controls the Systems They Run On

New research explains why static cyber insurance underwriting fails—and how AI-driven, continuous risk signals can improve pricing and prevention.

Why Static Cyber Insurance Models Are Failing in a World of Constant Risk

Function calling turned LLMs from chatbots into action systems—reshaping AI runtimes, security, reasoning models, and specialization.

AI in 2026: Function Calling, Reasoning Models, and a New Runtime Era

The shopping journey is shifting from clicks to AI-mediated conversations. Why today’s analytics frameworks aren’t built for what comes next.

Why Traditional Customer Journey Analytics Are Failing Modern Commerce

Learn how router/video/enterprise processes video over time—identifying objects, actions, and sequences.

OpenRouter’s Video Endpoint: The “Ask Your Video Anything” Model, Explained

UniVideo combines video comprehension, creation, and editing into a single system built on established components—here’s what it does and how to test it.

UniVideo Wants to Be the One Model That Understands, Generates, and Edits Video

Explore real-time facial segmentation with BiSeNet: how bisenet-faces parses faces into regions and how to tune speed vs accuracy for interactive apps.

The Model That “Understands” Your Face in Real Time

Explore probabilistic modeling on torical manifolds using von Mises, Wrapped Cauchy, and Kato-Jones distributions linked to Kuramoto models

Statistical Models on Circles and Tori: von Mises, Wrapped Cauchy, and Kato-Jones Distributions

<p><em>This is a Plain English Papers summary of a research paper called <a href="https://www.aimodels.fyi/papers/arxiv/step3-vl-10b-technical-report?utm_source=hackernoon&utm_medium=referral">STEP3-VL-10B Technical Report</a>. If you like these kinds of analysis, join <a href="https://www.aimodels.fyi/?utm_source=hackernoon&utm_medium=referral">AIModels.fyi</a> or follow us on <a href="https://x.com/aimodelsfyi">Twitter</a>.</em></p>

<h2>The efficiency paradox that rewrites multimodal AI</h2>
<p>Vision-language models face a cruel choice. Stay small and efficient, and you sacrifice reasoning capability. Scale to hundreds of billions of parameters, and you handle complex visual tasks effortlessly but burn enormous computational resources. The field has accepted this tradeoff as inevitable, like two ends of a rope you can only hold one end of at a time.</p>
<p>STEP3-VL-10B breaks this assumption by asking a different question: what if the bottleneck isn't model size at all, but how the model allocates its thinking?</p>
<p>A 10 billion parameter model that rivals systems 10 to 20 times larger sounds like marketing. The paper proves it's architecture. The approach combines three strategic choices made in sequence: unfrozen joint pre-training of vision and language components on 1.2 trillion multimodal tokens, over 1000 iterations of reinforcement learning that teaches the model to reason more efficiently, and a test-time reasoning method called Parallel Coordinated Reasoning that generates multiple interpretations of an image before synthesizing a final answer. The results are striking. STEP3-VL-10B reaches 92.2% on MMBench, 80.11% on MMMU, 94.43% on AIME2025, and 75.95% on MathVision, matching or exceeding models like GLM-4.6V-106B, Qwen3-VL-235B, and proprietary systems like Gemini 2.5 Pro.</p>
<p>But the real insight runs deeper than benchmark numbers. This work demonstrates that reasoning ability comes from multiple sources, and parameter count is only one of them. That realization changes what's possible for efficient AI.</p>
<h2>Teaching vision and language to grow up together</h2>
<p>Traditional vision-language models treat perception and language as separate problems that happen to live in the same system. A frozen Perception Encoder extracts visual features. A language model interprets those features. They've never learned to understand each other.</p>
<p>STEP3-VL-10B starts differently. During pre-training, both the Perception Encoder and the Qwen3-8B language decoder receive gradients simultaneously. They co-evolve across 1.2 trillion multimodal tokens. The Perception Encoder learns what patterns the language model actually cares about understanding. The language model learns what kinds of visual information matter for reasoning.</p>
<p>This sounds like a small change. It's not. Consider how humans learn language and visual understanding. You don't learn to recognize objects in isolation, then separately learn to talk about them. You learn them together, each informing the other. When you learn the word "bridge," you simultaneously understand what bridges look like, why they matter structurally, and how to reason about them. That integration is what STEP3-VL-10B captures at scale.</p>
<p>The scale matters too. 1.2 trillion tokens is large enough that both components can specialize meaningfully while still learning deep interdependencies. Too small, and emergent reasoning patterns never develop. Too large without the right architecture, and you're wasting compute. This number reflects the point where the two components have learned to cooperate fully.</p>
<p>The foundation established during pre-training is silent and invisible in benchmarks. You won't see a paper that says "unfrozen pre-training alone gave us 2% on MMBench." But without this step, the rest of the approach doesn't work. The model needs to enter RL training already understanding vision and language as integrated domains, not after-the-fact combinations.</p>
<h2>Learning to reason through structured iteration</h2>
<p>Pre-training establishes foundations. Reinforcement learning teaches strategy. The paper implements what it calls RLVR (Reinforcement Learning with Vision-Reasoning), running over 1000 iterations where the model generates multiple candidate reasoning chains for visual problems, receives rewards based on reasoning quality and correctness, and updates its weights to prefer better patterns.</p>
<p>What happens during this process is revealing. Figure 2 shows two curves that initially seem contradictory. The reward curve climbs steadily without saturating, indicating the model keeps finding room for improvement. Simultaneously, the average number of tokens generated initially spikes then decreases back toward baseline levels. This isn't a sign of failure or saturation. It's the opposite. The model is learning to reason more efficiently, achieving the same quality of reasoning in fewer tokens.</p>
<p><a href="https://arxiv.org/html/2601.09668/x4.png"><img src="https://arxiv.org/html/2601.09668/x4.png" alt="Performance trends showing reward increasing while token count decreases" /></a><br/><em>Performance trends showing reward increasing while token count decreases</em></p>
<p><em>RLVR dynamics: reward continuously improves (left) while the model learns to reason more concisely, reducing rollout tokens after initial exploration (right).</em></p>
<p>This pattern indicates genuine learning, not mere memorization or output inflation. The model isn't cheating by making longer outputs to seem smarter. It's developing better internal representations of visual reasoning. Early RL iterations teach basic structured thinking. Later iterations refine edge cases and teach multi-step integration across multiple hypotheses.</p>
<p>Figure 3 tracks this progress on representative benchmarks. Metrics on reasoning tasks (the kind where visual hypothesis exploration matters most) and perception tasks (the kind where all models converge to similar performance) both show the same pattern: rapid initial growth followed by steady improvement that mirrors the reward dynamics.</p>
<p><a href="https://arxiv.org/html/2601.09668/x5.png"><img src="https://arxiv.org/html/2601.09668/x5.png" alt="Benchmark performance over RL iterations" /></a><br/><em>Benchmark performance over RL iterations</em></p>
<p><em>Performance gains during RLVR track reward improvements: rapid initial growth followed by steady gains, indicating sustained learning rather than saturation.</em></p>
<p>The convergence between reward trends and benchmark trends confirms something important. The RL procedure isn't optimizing for gaming metrics or statistical artifacts. It's genuinely teaching the model better reasoning behavior, and that behavior transfers directly to standard benchmarks. This is crucial for believing that the approach will generalize beyond the specific tasks used during training.</p>
<h2>Why test-time scaling beats parameter scaling</h2>
<p>Here's where the conceptual breakthrough becomes concrete. Instead of scaling the model by making it larger, STEP3-VL-10B scales it at inference time through Parallel Coordinated Reasoning.</p>
<p>The mechanism is elegant. For a complex visual reasoning task, the model doesn't generate one answer. It generates multiple competing interpretations of the visual input. For a diagram, it might ask: what if the key relationship is spatial? What if it's conceptual? What if these elements are functionally connected rather than physically arranged? Each interpretation proceeds through independent reasoning chains. Then the model synthesizes across these chains, considering how different interpretations support or contradict each other, arriving at a final answer that integrates multiple hypotheses.</p>
<p>The term "Parallel Coordinated" captures both elements. The reasoning chains run in parallel, making efficient use of compute. But they coordinate, with the model tracking how different interpretations relate to each other. The final answer emerges from synthesis, not from selecting the loudest individual chain.</p>
<p>Figure 1 illustrates the consequence. STEP3-VL-10B bridges a gap that previous models couldn't cross. On pure perception tasks, where extracting visual information is straightforward, all models converge to similar performance. The curve flattens. But on complex reasoning tasks, where multiple interpretations are possible and need reconciliation, larger models normally pull away dramatically. STEP3-VL-10B with PaCoRe doesn't follow that pattern. Its performance on reasoning tasks rises alongside the perception baseline, breaking the historical correlation between reasoning difficulty and model size.</p>
<p><a href="https://arxiv.org/html/2601.09668/x3.png"><img src="https://arxiv.org/html/2601.09668/x3.png" alt="Comparison showing Step3-VL-10B bridging perception and reasoning gaps" /></a><br/><em>Comparison showing Step3-VL-10B bridging perception and reasoning gaps</em></p>
<p><em>STEP3-VL-10B with PaCoRe achieves frontier performance on both perception and reasoning tasks, bridging the traditional efficiency-capability gap.</em></p>
<p>This test-time scaling trades compute at inference for parameters in training. For many real-world applications, this is a profound advantage. You're spending inference tokens (generated at runtime) instead of training-time parameters. A batch processing system can spend more compute on a single request to get better accuracy. An interactive system can dial down test-time reasoning if latency becomes critical. An on-device application gets a model 10x smaller than the alternative while maintaining comparable reasoning capability. The flexibility is as valuable as the benchmark numbers.</p>
<p>This approach connects to broader work on test-time scaling across the field. OpenAI's o1 model and Anthropic's extended thinking both explore similar territory for language reasoning. STEP3-VL-10B demonstrates that the principle isn't language-specific. Multimodal systems can benefit equally from allocating more compute to test-time reasoning, particularly when visual interpretation ambiguity is involved.</p>
<h2>Where the model actually excels</h2>
<p>The benchmark results reveal something important: the approach's strengths are concentrated where reasoning matters most.</p>
<p>On MMBench (92.2%) and MMMU (80.11%), STEP3-VL-10B performs competitively but not dramatically ahead of the largest models. These benchmarks test perception and basic understanding. Multiple visual interpretations don't help much when the task is straightforward pattern recognition. The model shines on tasks where it needs to explore hypotheses.</p>
<p>Look at AIME2025 (94.43%), a competition mathematics benchmark that requires integrating visual information, problem setup, and multi-step reasoning. The model doesn't just extract numbers from diagrams; it interprets spatial relationships, reasons through constraints, and verifies answers against the original problem. MathVision (75.95%), focused specifically on visual mathematics and geometry, shows similar strength. These tasks demand the kind of hypothesis exploration that PaCoRe provides.</p>
<p>The split between perception and reasoning isn't accidental. It reveals the architecture's actual operating principle. When visual tasks are straightforward, you don't need multiple interpretations. When they require reasoning through ambiguity or multi-step inference, exploring alternative interpretations becomes powerful. The model learned to allocate its test-time reasoning where it's most valuable.</p>
<p>This specificity matters for practitioners considering deployment. STEP3-VL-10B is exceptional for applications with complex visual reasoning: mathematics and geometry, engineering diagram analysis, scientific image interpretation, medical imaging in domains requiring reasoning about relationships rather than classification. For tasks where perception dominates and reasoning is minimal, a smaller model might suffice.</p>
<h2>What the approach can't do (yet)</h2>
<p>Every elegant solution has boundaries. Understanding them prevents misuse and clarifies where different approaches remain necessary.</p>
<p>Latency becomes a constraint when test-time reasoning is needed. Parallel Coordinated Reasoning works by generating multiple reasoning paths. A system requiring sub-100-millisecond response times can't afford this exploration. Real-time interactive applications, mobile interfaces with strict latency budgets, and streaming systems where decisions must be made instantly need different architectures or models optimized for speed over reasoning quality.</p>
<p>The language backbone, borrowed from Qwen3-8B, becomes a limitation for tasks with minimal visual content or requiring extensive reasoning over pure language. A problem involving a single small diagram and 50 pages of text might not benefit from STEP3-VL-10B's strengths. Large language models with stronger language modeling components would perform better. The model was optimized for vision-language integration, not language-heavy tasks.</p>
<p>Domain specialization matters too. The training emphasized visual reasoning across diverse domains. Specialized fields requiring deeply specific language knowledge (legal document understanding with minimal images, scientific text with citation-heavy reasoning, domain-specific jargon) might see better results from models trained specifically for those domains. STEP3-VL-10B's strength is breadth; specialized applications might trade some breadth for depth elsewhere.</p>
<p>These limitations aren't failures. They're the natural boundaries of an approach optimized for specific tradeoffs. A smaller, faster model for latency-critical applications. A larger language backbone for language-heavy reasoning. These coexist with STEP3-VL-10B in a broader toolkit rather than being replaced by it.</p>
<h2>The research reproducibility layer</h2>
<p>The paper releases the full model suite open-source. This is more than a nice addition to the research narrative; it's a statement about how knowledge should move in fast-moving fields.</p>
<p>Open release means researchers can use STEP3-VL-10B directly, fine-tune it for specific domains, and study what actually drives its performance rather than inferring from published claims. Training details sufficient for reproduction are included, though not every step-by-step training configuration. The benchmark evaluation procedures are documented clearly enough that future work can claim improvements with confidence rather than uncertainty about whether they used the same setup.</p>
<p>This matters because multimodal AI is moving fast. When models are proprietary and benchmarked behind those companies' methodologies, research gets gated. One group's research direction becomes the constraint for a thousand other researchers trying to build on it. Open models and reproducible baselines accelerate progress across the field. They also build trust in claims. When someone says "our new method beats STEP3-VL-10B," the community can verify whether that's true, whether they're testing the same way, and whether the comparison is fair.</p>
<p>The release also matters for practitioners. A frontier-performing open model at 10 billion parameters is genuinely valuable for on-device inference, cost-effective deployment, and experimentation in resource-constrained settings. Organizations can use this as a baseline, fine-tune it for their domains, and maintain control over their systems rather than relying on proprietary APIs.</p>
<h2>Reframing how multimodal AI scales</h2>
<p>STEP3-VL-10B is a specific model. It's also a proof of concept that challenges how the field thinks about scaling multimodal systems.</p>
<p>The traditional frame went like this: larger model equals smarter model, so efficiency means making models smaller while accepting capability reductions. The new frame this work enables is more nuanced. Reasoning capability comes from four sources: parameter count, pre-training scale, post-training refinement through RL, and test-time compute allocation. The optimal mix depends on the application. A research system where latency doesn't matter might push hard on test-time compute. A mobile application might minimize both parameter count and test-time reasoning. An on-device system might increase parameter count slightly to reduce test-time overhead.</p>
<p>This shift isn't theoretical. Major AI labs are already exploring test-time scaling for language reasoning. OpenAI's o1 model represents a massive commitment to test-time compute. Anthropic's extended thinking explores similar scaling. STEP3-VL-10B shows this principle works equally well for multimodal systems, where visual ambiguity and interpretation complexity create natural opportunities for hypothesis exploration.</p>
<p>The practical implications cascade outward. Future multimodal systems might look less like "bigger models" and more like "smarter inference procedures." A 10 billion parameter model becomes competitive with 100 billion parameter models through better reasoning allocation, not through different training data or longer training. This opens possibilities for cost-effective scaling, edge deployment that maintains frontier reasoning capability, and flexible systems where you can trade inference cost for quality per request.</p>
<p>The field is learning that the frontier of multimodal AI doesn't need to be a frontier of model size. It can be a frontier of architectural innovation, test-time reasoning, and thoughtful resource allocation. STEP3-VL-10B demonstrates that this isn't a small efficiency gain at the margins. It's a fundamental rethinking of how to build capable systems that competes directly with massive parameter-scaling approaches.</p>

<hr/><p><strong>Original post:</strong> <a href="https://www.aimodels.fyi/papers/arxiv/step3-vl-10b-technical-report?utm_source=hackernoon&utm_medium=referral">Read on AIModels.fyi</a></p>

Breaking the Perception–Reasoning Tradeoff in Vision-Language Models

Agentic reasoning reframes LLMs as autonomous agents that plan, act, and learn—unlocking real-world applications beyond static text prediction.

Notes on Agentic Reasoning for Large Language Models

Kling v2.6 Motion Control turns reference materials into directed performances—ideal for creators, cutscenes, and virtual influencers whoneed consistent motion.

Kling v2.6 Motion Control: The “Do Exactly This” Text-to-Video Model

MODE is a reliable data collection solution for real-world data

MODE

tinkermode.com

11-50 emps

Since 2014

MODE's stories on HackerNoon