OpenCode Go Models

Newest model in the lineup. MiniMax Sparse Attention (MSA) makes 1M-token context practical — 1/20th per-token compute vs prior gen, 9× faster prefill, 15× faster decode. Native text+image+video input. Open weights promised within days of launch.

Best SWE-Bench Pro score in the Go lineup (60.6%). Flagship agent model — ran a 35-hour autonomous kernel-optimization session: 1,158 tool calls, 432 evaluations, 10x speedup, zero human intervention. Highest-ranked Chinese model on the AA Intelligence Index (56.6).

Native multimodal agentic flagship. 1T MoE with ~32B active; MLA keeps long-context VRAM cheap. Agent-swarm capable: scales to 300 sub-agents executing 4,000 coordinated steps. Beats Opus 4.6 and GPT-5.4 on SWE-Bench Pro.

Biggest model in the lineup: 1.6T params, 49B active. CSA+HCA hybrid attention — at 1M context it needs only 27% of the FLOPs and 10% of the KV cache of V3.2. Pro-Max variant posts the highest coding scores of any model (80.6% SWE-Verified, 93.5 LiveCodeBench).

The throughput king of Go — 31,650 requests per 5 hours, the highest allowance in the plan. Efficiency-optimized MoE (284B total, just 13B active) with the same 1M-token context as Pro. The default pick for high-volume agent loops.

Replaces both MiMo-V2-Pro and V2-Omni: one open-source model for reasoning AND full multimodality (image, audio, video, text). 1.02T MoE, hybrid attention + 3-layer multi-token prediction. Uses 40–60% fewer tokens than Opus 4.6 for comparable agentic scores. UltraSpeed variant hits 1,000 tok/s.

The efficient sibling in Xiaomi's V2.5 public-beta series (which also includes TTS and ASR variants). Second-highest allowance on Go at 30,100 requests per 5 hours — built to be hammered in agent loops where Pro-class reasoning isn't needed.

Multimodal variant of the 3.7 series with vision input at a lower price point. The budget pick for high-volume, routine workloads — Max is the heavy hitter, Plus is what you run all day. Generous 4,300 req/5h on Go.

Prior-generation Plus, kept in the lineup as a proven high-volume option (3,300 req/5h). Solid general coding and chat at minimal cost — the safe fallback when 3.7 capacity is constrained.

Long-horizon agentic flagship. Can work autonomously for up to 8 hours on a single task. +20.4pts CyberGym, +7.3pts Terminal-Bench vs GLM-5. Trained entirely on Huawei Ascend chips.

First open-weights model to break AA Intelligence Index 50. DeepSeek Sparse Attention for 200K context. Superseded by 5.1 for agentic work but still a strong general reasoner.

Native multimodal agentic model built via continual pretraining on ~15T visual+text tokens atop K2. MoonViT 400M vision encoder. State-of-the-art visual coding and BrowseComp w/ ctx (74.9%).

First model deeply participating in its own evolution. Self-improving with Agent Teams and dynamic tool search. Highest GDPval-AA Elo (1495) among open-source models. 97% skill adherence.

Open-source workhorse with best-in-class SWE-Bench Verified (80.2%) at launch. ~100 tok/s inference. Matching Opus-class coding at 1/20th the cost — still the value benchmark of the lineup.