This guide examines six open-weight reasoning models trained with RL, alongside Modal, the AI infrastructure platform purpose-built for deploying and training these models at scale. Whether you're running inference on a 685B-parameter model or fine-tuning smaller reasoning models for specialized tasks, the combination of the right model and the right infrastructure determines your success.
Key Takeaways
- Reinforcement learning drives reasoning breakthroughs: RL-trained models like DeepSeek-V3.2-Speciale report 96.0% on AIME 2025, above the GPT-5-High and Gemini-3.0-Pro results listed under the paper's evaluation protocol
- Model size doesn't determine reasoning capability: Microsoft's Phi-4-reasoning-plus achieves 81.3% on AIME 2024 with just 14B parameters, matching or outperforming some much larger models on mathematical reasoning
- Infrastructure matters as much as model selection: Deploying 235B+ parameter reasoning models requires GPU clusters with high-bandwidth interconnects and efficient memory management
- Hybrid thinking modes enable flexible deployment: Qwen3-235B-A22B supports both rapid responses and deep chain-of-thought reasoning in a single unified model
- Permissive licensing supports commercial use: MIT and Apache 2.0 generally permit commercial use and fine-tuning, subject to their notice, attribution, patent, and other applicable legal requirements
- Modal provides purpose-built infrastructure: With multi-node GPU clusters featuring B200/H200/H100 GPUs and up to 3,200 Gbps of RDMA networking, Modal handles the infrastructure complexity so teams can focus on model deployment
1. Modal
Modal delivers the AI infrastructure that developers need to deploy, train, and scale open-weight reasoning models without managing GPU clusters or configuring distributed systems. While the other entries in this list are reasoning models, Modal is the platform that makes running them practical at production scale.
How Does Modal Work?
Modal's serverless compute platform takes your code, packages it into containers, and executes it in the cloud with automatic scaling. Key capabilities include:
- Fast cold starts: Engineered for fast cold starts and faster feedback loops, with an optimized filesystem that helps containers come online quickly without letting large images slow startup down
- Autoscaling: Scale from zero toward 1,000+ GPUs in minutes based on demand, paying only for actual compute time
- Multi-cloud capacity pool: Access B200, H200, H100, A100, and L40S GPUs across a deep multi-cloud capacity pool without managing hyperscaler reservations or cloud-provider quotas
- Memory snapshotting: CPU Memory Snapshots, and Alpha GPU Memory Snapshots, can reduce initialization-heavy cold starts by restoring previously initialized CPU or GPU state
Why Modal for Reasoning Models
Reasoning models present unique infrastructure challenges. A 685B-parameter model like DeepSeek-V3.2-Speciale requires significant GPU memory and compute. Modal's platform primitives address these challenges:
- High-performance inference: Modal's optimized GPU scheduler and inference infrastructure deliver low-latency responses for reasoning workloads
- Multi-node training (Beta): Train or fine-tune reasoning models across GPU clusters with up to 3,200 Gbps of RDMA scale-out networking
- Persistent storage: Volumes store model weights efficiently, eliminating repeated downloads
- Secure sandboxes: Run AI-generated code safely in isolated containers at massive concurrency, in whatever language or runtime the workload requires
Enterprise-Ready Platform
Modal powers cloud infrastructure for over 10,000 teams. Modal has successfully completed a SOC 2 Type II audit and is SOC 2 Type II compliant, with HIPAA support for Enterprise customers via a signed BAA. Enterprise customers can purchase Modal through AWS or Google Cloud Marketplace and apply eligible committed cloud spend to Modal usage.
Best For: Teams deploying open-weight reasoning models who need production-grade GPU infrastructure with fast container cold starts, autoscaling to 1,000+ GPUs in minutes, and pay-per-second billing.
2. DeepSeek-V3.2-Speciale
DeepSeek-V3.2-Speciale was one of the strongest open-weight reasoning releases of late 2025, particularly on the benchmarks reported in its technical paper. The model reported gold-medal-level performance on the 2025 IMO and IOI competitions, a milestone for open-weight models.
Architecture and Training
DeepSeek-V3.2-Speciale uses a Mixture of Experts (MoE) architecture with 685B total parameters and 37B activated per token. This design delivers frontier-level reasoning while maintaining practical inference costs. The model was trained with extensive reinforcement learning using reduced length penalties. In DeepSeek's reported evaluations, average output length varied substantially by benchmark, from roughly 23K tokens on AIME 2025 to about 77K tokens on Codeforces.
Benchmark Performance
The model's reasoning capabilities are documented across multiple olympiad-level benchmarks:
- AIME 2025: 96.0% accuracy, above the GPT-5-High result of 94.6% listed in the paper
- IMO 2025: 35/42 points, achieving gold-medal level
- IOI 2025: 492/600 points in competitive programming
- ICPC World Finals: 10/12 problems solved
- Codeforces Rating: 2701 (Grandmaster tier)
Technical Innovations
- DeepSeek Sparse Attention (DSA): An attention mechanism that substantially reduces the dominant long-context attention cost through sparse token selection, although its lightweight indexer retains a quadratic component
- Extended reasoning chains: Trained to produce longer reasoning traces, generating comprehensive step-by-step solutions
- 128K context window: Handles complex multi-step problems that require extensive context
Best For: Research teams and organizations requiring very high reasoning accuracy for mathematical, scientific, and programming applications. Deploy on Modal's GPU infrastructure for production workloads.
3. Qwen3-235B-A22B
Qwen3-235B-A22B brings a distinctive capability to open-weight reasoning: hybrid thinking modes that let you control reasoning depth per request. This flexibility makes it a strong option for production deployment.
Hybrid Thinking Architecture
Qwen3 supports both rapid responses and deep chain-of-thought reasoning in a single model deployment. Users can toggle between modes using /think and /no_think flags, and even specify token budgets for reasoning. The model naturally adapts to incomplete thinking scenarios, making it practical for latency-sensitive applications.
Training Methodology
The model underwent multi-stage post-training including:
- Long chain-of-thought cold start training
- Reasoning-focused reinforcement learning
- Thinking mode fusion for unified deployment
- General RL for broad capability
Key Specifications
- Parameters: 235B total, 22B activated per token
- Context Window: 32,768 tokens natively; up to 131,072 tokens with YaRN
- Languages: 119 languages and dialects according to the Qwen3 technical report
- License: Apache 2.0
Benchmark Results
- AIME 2024: 85.7%
- AIME 2025: 81.5%
- LiveCodeBench v5: 70.7%
- Codeforces: 2,056 rating
In the Qwen3 report's base-model comparison, Qwen3-235B-A22B-Base outperformed DeepSeek-V3-Base on 14 of 15 listed benchmarks while using only one-third the total parameters and two-thirds the activated parameters.
Best For: Production teams needing flexible reasoning depth control without model switching. Use Modal's batch processing for high-throughput inference workloads.
4. GPT-OSS-120B
GPT-OSS-120B marked OpenAI's return to releasing open-weight language models. This open-weight release brought competitive reasoning quality, which OpenAI reports approaches or matches some of its own models on selected evaluations, in a model that fits on a single 80 GB GPU.
Model Specifications
- Parameters: 117B total, roughly 5.1B activated per token (MoE architecture)
- Context Window: 128K tokens
- License: Apache 2.0
- Release: August 5, 2025
Performance Metrics
- AA-IFBench: 69.0% on instruction following, per BenchLM's July 2026 leaderboard (a third-party evaluation rather than an OpenAI-reported result)
- Deployment footprint: Fits on a single 80 GB GPU, per OpenAI's launch material
Why It Matters
GPT-OSS-120B reflects OpenAI's participation in the open-weight ecosystem alongside its proprietary offerings. The model uses OpenAI's Harmony Response Format, an OpenAI-specific format (released under Apache 2.0) that defines the conversation, reasoning, and function-call structure GPT-OSS expects for correct operation.
Best For: Organizations wanting a permissively licensed open-weight model with strong reported reasoning results for commercial deployment.
5. GLM-4.5
GLM-4.5 from Zhipu AI (Z.ai) is an RL-trained, hybrid-thinking MoE model with a 128K context window, well suited to long-form code and document generation. A larger context window can reduce the need to split big inputs across multiple requests, although complex code-generation workflows generally remain iterative.
Architecture Highlights
- Parameters: 355B total, 32B activated per token
- Context Window: 128K tokens
- License: MIT
Technical Innovations
- Hybrid thinking MoE: Trained with reinforcement learning for reasoning, coding, and agentic tasks, with 32B parameters active per token out of 355B total
- 128K context window: Handles long inputs and outputs within the combined sequence budget for code and document workflows
Benchmark Performance
- SWE-Bench Verified: 64.2% in its technical report
- MMLU-Pro: 84.6 under Z.ai's launch evaluation
GLM-4.5 is well suited to autonomous code generation and documentation synthesis tasks. Note that GLM-4.5 has since been followed by later GLM generations; it is included here as a widely deployed, MIT-licensed model with a manageable hardware footprint and reproducible published results.
Best For: Development teams working on long-context code synthesis, documentation, or technical writing applications.
6. MiniMax-M1-80k
MiniMax-M1-80k uses large-scale reinforcement learning over reasoning, coding, tool-use, and software-engineering tasks, with an emphasis on efficient long test-time computation.
Training Approach
MiniMax-M1 was trained with large-scale reinforcement learning across reasoning, coding, tool-use, and software-engineering problems. Its reported RL data includes roughly 30,000 coding samples, several thousand software-engineering samples, and about 25,000 general RL samples, along with logic tasks and sandbox-based software-engineering environments. The design emphasizes efficient long test-time computation rather than benchmark scores alone.
Model Specifications
- Parameters: 456B total, 45.9B activated per token
- Context Window: 1M tokens native; M1-80k is the variant trained for an 80K reasoning/output budget
- License: Apache 2.0
Performance Characteristics
- MMLU-Pro: 81.1% in MiniMax's reported evaluation
- AA-IFBench: 41.8% on BenchLM's July 2026 leaderboard (a third-party evaluation)
- Efficient long-sequence computation: Its hybrid/lightning-attention design reduces computational cost for long reasoning sequences (for example, lower attention FLOPs than DeepSeek-R1 at a 100K generation length); this is not equivalent to using fewer reasoning tokens
MiniMax-M1 activates approximately 45.9B of its 456B parameters per token, or about 10.1%, and is designed to scale test-time computation efficiently through 40K- and 80K-thinking variants. Note that MiniMax has since released later generations; M1 is included here for its long-context reasoning and its permissively licensed, reproducible published results rather than as MiniMax's current agentic model.
Best For: Teams building agentic AI applications that require practical task completion and long-context reasoning. Deploy agents using Modal's secure sandboxes for isolated code execution.
7. Phi-4-reasoning-plus
Microsoft's Phi-4-reasoning-plus proves that advanced reasoning doesn't require massive parameter counts. At just 14B parameters, this dense model outperforms much larger alternatives on mathematical reasoning tasks.
Small Model, Big Reasoning
At 14B parameters, Phi-4-reasoning-plus has roughly one-seventeenth the parameters of a 235B model like Qwen3-235B-A22B. Microsoft accomplished this through:
- High-quality synthetic datasets for training
- Organic data curation from diverse sources
- Post-training innovations specific to reasoning tasks, including supervised reasoning fine-tuning and additional reinforcement learning
Benchmark Performance
The "plus" variant leads on 9 of 11 shared benchmarks compared to standard Phi-4-reasoning:
- AIME 2024: 81.3% vs 75.3% for standard variant
- OmniMath: 81.9% vs 76.6%
- FlenQA: 97.9%
- GPQA: 68.9%
- IFEval: 84.9%
Practical Advantages
- 14B dense architecture: Can run on a single sufficiently large GPU; 24 GB-class cards generally require quantization, reduced context, or offloading
- MIT license: Full commercial use permitted
- Efficient footprint: Lower memory and hardware requirements than the large MoE models in this list
Microsoft's research demonstrates that data quality and post-training innovation can achieve competitive reasoning performance without scaling to hundreds of billions of parameters.
Best For: Teams needing capable reasoning on constrained hardware or those wanting to run inference locally. For cloud deployment, Modal's GPU infrastructure provides efficient single-GPU instances.
Why Modal Stands Out for Reasoning Model Deployment
Purpose-Built for AI Workloads
Modal's infrastructure is engineered specifically for AI workloads like reasoning model inference and training. The platform's custom container runtime, scheduler, and file system optimize every layer for GPU-intensive operations.
Production-Ready GPU Infrastructure
Deploying a 685B-parameter reasoning model requires more than raw GPU access. Modal provides:
- B200, H200, and H100 availability: Access recent GPU generations without managing cloud-provider reservations
- Private cluster networking: Up to 3,200 Gbps RDMA interconnects for multi-node training and inference
- Memory snapshotting: CPU Memory Snapshots, and Alpha GPU Memory Snapshots, reduce initialization-heavy cold start latency by restoring previously initialized state
- Regional routing (Beta): Modal runs a global GPU fleet and lets you select a container region or set a routing region closer to your users; standard Function traffic defaults to us-east
Developer Experience
Modal's code-first SDKs, available in Python, TypeScript, and Go, make deploying reasoning models feel local while running in the cloud. Define your container image, specify GPU requirements, and deploy, and the platform handles scaling, networking, and orchestration.
Flexible Deployment Patterns
Reasoning models serve diverse use cases, and Modal supports them all across CPU and GPU compute:
Enterprise Security and Compliance
Modal has successfully completed a SOC 2 Type II audit and is SOC 2 Type II compliant, and offers HIPAA support (via BAA for Enterprise customers), enabling deployment in regulated environments. The platform uses gVisor-based isolation for compute jobs, with TLS 1.3 on all public Modal APIs and user data encrypted in transit and at rest.
Proven at Scale
Modal powers infrastructure for over 10,000 teams running inference, training, and batch processing workloads. The platform's multi-cloud capacity pool improves GPU availability during demand spikes and reduces reliance on any single provider.
For teams deploying open-weight reasoning models in production, Modal reduces infrastructure complexity while delivering the performance these demanding workloads require.
Deploy your first reasoning model on Modal's GPU infrastructure.
Get Started Free