What Granite-3.3-8B-Math-PRM-v2 is
Granite-3.3-8B-Math-PRM-v2 is an 8-billion-parameter generative process reward model from IBM Research’s Granite Alignment Team. A process reward model, often abbreviated PRM, scores the intermediate steps of a solution rather than judging only the final result. This distinction matters because a final answer can be correct by chance, while a step-level evaluator can identify where a reasoning chain first becomes unreliable.
The model is fine-tuned from Granite-3.3-8B-Instruct and is listed as a current open-weight experimental model in IBM Granite’s catalog. It is available as a downloadable model repository rather than as a first-party hosted service with published token pricing. The canonical repository identifier is ibm-granite/granite-3.3-8b-math-prm-v2.
Its documented release date is January 1, 2026, and it is distributed under the Apache 2.0 license. Those details make it suitable for teams that want to inspect, deploy, and integrate the model in their own reasoning systems, subject to their own infrastructure and compliance requirements.
How the model evaluates reasoning
Granite-3.3-8B-Math-PRM-v2 expects a mathematical problem and a response divided into reasoning steps. At the end of each step, the inference prompt asks: Is this response correct so far (Y/N)? The model then generates a Y or N judgment.
The generated letter is useful, but the probability assigned to Y is often more valuable in an automated pipeline. That probability can be treated as a numerical reward score. An application can score every step in several candidate solutions, combine the scores, and select the candidate with the strongest evidence of a correct reasoning path.
For example, a solver might produce 16 possible solutions to an algebra problem. Instead of choosing the most common final answer, a controller can ask the PRM to assess each intermediate step, penalize candidates that introduce an invalid transformation, and select or rerank the remaining solutions. The same approach can support weighted voting, error localization, and training or evaluation workflows based on process supervision.
Primary use cases
- Best-of-N selection: generate multiple solutions and choose the candidate whose reasoning receives the strongest step-level scores.
- Mathematical error detection: identify the point at which a solution appears to depart from a correct argument.
- Process supervision: provide feedback about intermediate reasoning rather than relying only on a final-answer label.
- Weighted voting: give more influence to candidate solutions whose steps receive higher correctness probabilities.
- Reasoning-chain evaluation: compare solution strategies and measure how reliably they maintain correctness from one step to the next.
IBM also reports inference-scaling results on coding benchmarks. This suggests that the model may be useful in coding-related evaluation experiments when several candidate solutions are generated and checked. However, it should not be treated as a general coding assistant: its core task remains process evaluation, not code production.
Strengths and reported evaluation
The model’s main strength is specialization. A conventional language model may be able to produce a mathematical answer, but Granite-3.3-8B-Math-PRM-v2 is specifically configured to assess whether the reasoning is correct so far. This makes it a natural component for systems that separate generation from verification.
IBM reports strong best-of-N results on Math-500, including improvements over majority voting across several candidate counts. The model was also evaluated on ProcessBench and PRMBench. Reported PRMBench figures include an overall score of 65.1 and an average sensitivity score of 75.4. These are process-evaluation results; they should not be read as general language-model quality scores or as evidence that the model is intended to replace a chat assistant.
The research data assigns an editorial reasoning score of 9 out of 10 and a coding score of 7 out of 10. These are estimates based on the model’s specialized role and published evaluations, not IBM-published ratings. The model’s reasoning score reflects its suitability for checking mathematical chains, while the coding score reflects the more limited evidence for coding-related inference scaling rather than code-generation ability.
Technical specifications and supported modalities
| Specification | Available information |
|---|---|
| Provider | IBM |
| Model family | Granite 3.3 |
| Parameters | 8 billion |
| Model type | Reasoning and process reward model |
| Input | Text |
| Output | Text, primarily Y/N judgments and associated probabilities |
| Context length | 131,072 tokens, based on Granite 3.3 8B base/instruct configuration lineage |
| Maximum output tokens | Not separately published for this exact PRM |
| License | Apache 2.0 |
| Hosted token pricing | Not published for this exact model |
The model accepts text and produces text. It is not documented as supporting image, audio, or video input or output. It also is not documented as a tool-calling or function-execution model. Streaming is listed in the supplied catalog data, but because this is a downloadable open-weight model, the practical streaming behavior depends on the serving framework and deployment configuration.
The 131,072-token context figure should be interpreted carefully. It comes from the Granite 3.3 8B base/instruct configuration lineage rather than a separately stated PRM-specific limit in the model card. The exact model card also does not publish a maximum generated-output length. Users should therefore verify the effective context and generation settings in their chosen Transformers, vLLM, or SGLang deployment.
Deployment and pricing
Granite-3.3-8B-Math-PRM-v2 is released as downloadable open weights. The model card provides Transformers-based usage information and identifies local deployment options such as vLLM and SGLang. This gives an engineering team control over hosting, batching, prompt formatting, and the way step scores are aggregated.
There is no verified IBM first-party hosted endpoint or published per-token input and output price for this exact model in the supplied information. Consequently, there is no reliable hosted price to compare with commercial API models. The real cost of using it depends on the hardware, serving stack, utilization, and number of candidate solutions scored. A PRM can also increase total inference work because a system may generate and evaluate multiple candidates rather than making one model call.
Its editorial cost score is 7 out of 10, but that score is not a vendor price claim. It reflects the potential cost advantage of open-weight deployment and the lack of per-token licensing information, balanced against the infrastructure required to run an 8-billion-parameter model.
Limitations and trade-offs
The most important limitation is that this is not a polished-answer model. Its intended output is a correctness signal, usually a Y/N judgment or the probability associated with Y. An application that needs explanations, conversational responses, or a final mathematical solution will need a separate generator or an additional processing layer.
Prompt formatting and step segmentation are also important. The model is designed to judge each step in sequence, so poorly separated reasoning, ambiguous transitions, or an incorrectly placed evaluation prompt may produce less useful scores. Applications may need to calibrate the probabilities, decide how to combine step scores, and determine how to handle a single low-confidence step.
There is no separately documented knowledge-cutoff date for this exact process reward model. Its specialized evaluation role also means that benchmark performance should not be generalized to unrelated tasks. Although the underlying Granite base family supports multiple languages, the supplied model information identifies English as the fine-tuning target for this PRM.
When to choose Granite-3.3-8B-Math-PRM-v2
Choose this model when the central problem is deciding whether mathematical reasoning is correct at the step level. It is especially appropriate for a self-hosted best-of-N system, a research project on verifier-guided reasoning, a benchmark harness, or a pipeline that needs to identify errors before accepting a generated solution.
Its open-weight Apache 2.0 distribution is useful when deployment control matters more than turnkey API access. Teams can run it beside a separate generator and design their own scoring, reranking, and logging logic. The model is also a reasonable choice when the system needs a specialized verifier rather than another general-purpose assistant.
Choose another type of model when the requirement is direct user conversation, final-answer generation, multimodal understanding, speech, image generation, tool use, or a simple hosted API with predictable per-token billing. A general-purpose instruct model is more appropriate for producing explanations, while a dedicated code-generation model is better suited to writing software. In those cases, Granite-3.3-8B-Math-PRM-v2 may still serve as a secondary evaluator, but it should not be the only model in the workflow.
Bottom line
Granite-3.3-8B-Math-PRM-v2 occupies a narrow but useful role in IBM’s Granite lineup. It is an open-weight mathematical process reward model for checking reasoning chains, not a general chat system. Its Y/N judgments and reward probabilities can improve candidate selection and expose intermediate errors, particularly when paired with a separate solution generator. The trade-off is additional deployment and orchestration work, no published hosted price, and no guarantee that it will produce the kind of final response expected from an assistant model.