1. Introduction
GRM-2.6-Plus-0628 is a 27B-parameter reasoning model and a small update to GRM-2.6-Plus, built for general-purpose AI and optimized for difficult, high-complexity tasks. It is designed to deliver stronger performance for its size while remaining practical, efficient, and accessible for advanced local and research-oriented use.
This version improves upon GRM-2.6-Plus with a focus on long-horizon agentic tasks and the ability to solve harder problems, allowing it to better compete head-to-head with frontier models. The model focuses on structured reasoning, helping it produce more accurate, coherent, and reliable responses across demanding problems. GRM-2.6-Plus-0628 brings elite-level reasoning to complex workloads, making it suitable for users who need a capable model for advanced problem-solving, coding, agents, and everyday intelligence.
2. Key Capabilities
- Elite-Level Reasoning for Hard Tasks: GRM-2.6-Plus-0628 is optimized to handle difficult reasoning workloads with clarity, consistency, and strong step-by-step problem-solving ability.
- Improved Long-Horizon Agentic Performance: This update specifically targets long-horizon agentic workflows, enabling the model to maintain coherence and effectiveness across extended multi-step tasks.
- High Performance for Its Size: With 27B parameters, the model is designed to deliver excellent capability relative to its scale, balancing strong intelligence with practical deployment.
- Advanced Coding and Agentic Use: GRM-2.6-Plus-0628 is well suited for code generation, structured problem-solving, tool-style workflows, and local agentic applications.
- Optimized for Practical Deployment: The model aims to remain efficient and usable across capable consumer and workstation hardware while offering strong performance for advanced tasks.
[!Important] 2026-07-27 Update — Chat template fix (tool calls inside reasoning), v2.
Fixes tool calls landing in
reasoning_contentinstead oftool_callson llama.cpp and other engines that rely on the template's forced-open<think>prefill.(1) Explicit close directive: the
<IMPORTANT>block now requires</think>to be closed in every response, unconditionally, with no conversational text before the tool call.(2) Closing-tag variant detection: the history parser now recognizes
</think>,</thinking>, and common malformed variants (</ think>,</think >), not just the exact tag — borrowed from community testing on the wider Qwen3.5/3.6 template ecosystem.(3) Empty-think guard: past turns with no actual reasoning content no longer render an empty
<think>\n\n</think>block. Empty-think injection was training the model, in-context, to associate "no thinking" with "tool call," which made the bug worse over multi-turn agentic loops.(4) History parse fallback: assistant turns that contain
<tool_call>but no closing think tag are split at the tool-call boundary instead of dumping the raw XML as plain text into content.(5)
auto_disable_thinking_with_toolskwarg (optional, defaultfalse): scoped to individual turns that follow atool_response, not session-wide — so planned next-steps in an agentic loop skip the forced<think>block, while turns that genuinely need extended reasoning keep it.No action needed beyond replacing
chat_template.jinja. Upstream tracking:ggml-org/llama.cpp#22684,#20260.
3. Performance
GRM-2.6-Plus-0628 is designed to be a highly capable 27B local AI model for complex reasoning, coding, everyday chat, and agentic workflows. It focuses on delivering better performance for its size, making it a strong option for users who want powerful reasoning without relying only on massive-scale models.
Its core strength is practical intelligence: elite-level reasoning, strong task understanding, stable responses, and the ability to handle difficult problems across multiple domains.
Detailed Benchmarks
| GRM-2.6-Plus-0628 | GRM-2.6-Plus | Qwen3.6-27B | google/gemma-4-31B-it | GPT-5.4-Mini | Claude-4.5-Haiku | |
|---|---|---|---|---|---|---|
| Knowledge & STEM | ||||||
| MMLU-Pro | 88.1 | 86.8 | 86.2 | 85.2 | -- | 80.0 |
| MMLU-Redux | 96.4 | 94.2 | 93.5 | 93.7 | -- | -- |
| C-Eval | 92.4 | 92.0 | 91.4 | 82.6 | -- | -- |
| GPQA Diamond | 90.1 | 88.3 | 87.8 | 84.3 | 88.0 | 73.0 |
| SuperGPQA | 67.5 | 66.4 | 66.0 | 65.7 | -- | -- |
| Reasoning & Coding | ||||||
| LiveCodeBench v6 | 86.5 | 84.8 | 83.9 | 80.0 | -- | 51.1 |
| HMMT Feb 26 | 85.9 | 84.8 | 84.3 | 77.2 | -- | -- |
| AIME26 | 95.6 | 95.1 | 94.1 | 89.2 | -- | -- |
| General Agent | ||||||
| SWE-bench Verified | 79.7 | 77.7 | 77.2 | 52.0 | -- | 73.3 |
| SWE-bench Pro | 56.1 | 54.0 | 53.5 | 35.7 | 54.4 | -- |
| Terminal-Bench 2.0 | 62.6 | 59.8 | 59.3 | 42.9 | 60.0 | 41.0 |
4. Family
The GRM-2.6 family is available in various sizes to suit every case.
| Model | Size | Domain |
|---|---|---|
| GRM-2.6-Plus-0628 | 27B | Updated model for extremely difficult tasks with improved long-horizon agentic performance |
| GRM-2.6-Plus | 27B | Powerful model for extremely difficult tasks |
| GRM-2.6-Opus | 27B | Merge of GRM-2.6-Plus optimized for difficult terminal and coding tasks |
5. Architecture
GRM-2.6-Plus-0628 is built on the Qwen3.6 architecture and is optimized for complex tasks, agent environments, and everyday chat.
GRM-2.6-Plus-0628 applies the same principle to a stronger, larger foundation, resulting in a model that punches above its weight class on structured reasoning tasks while remaining deployable on consumer hardware.
GRM-2.6-Plus-0628 is developed by OrionLLM and released under the Apache 2.0 License.