Sage-A1.5's headline R9700 result depends heavily on how it is prompted to reason. With thinking disabled, it reached 86.75% on MMLU-Pro, 98% on GSM8K, and 6/6 coding tasks. In a separate greedy run with thinking enabled, it looped on self-verification and fell to 55.25%, 26%, and 0/6. These are results from different inference modes, not evidence that the model is uniformly strong or weak.
About the model
InternScience's Sage-A1.5 model card identifies the model as a 35B mixture-of-experts system built on Qwen3.6 and Agents-A1 from Shanghai AI Laboratory. The name and card describe a 35B-A3B class model, with about 3 billion active parameters per token. It accepts image and text input and is published under Apache 2.0. The card documents a 262,144-token context window and provides serving examples for both ordinary and MTP-enabled inference.
InternScience positions Sage as an agentic model for long-horizon scientific problem solving, with adaptive orchestration, verification, and research discovery. The card reports strong results on its own tool-use and research-oriented evaluations, including GAIA, SWE-bench Pro, and science benchmarks. Those are the publisher's reported evaluations and should be kept distinct from our local suite. The model card recommends temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 1.5, and repetition penalty 1.0. It also says its orchestration and verification modes are opt-in features, separate from the basic model weights.
What we measured
We tested Sage-A1.5 Q4_K_M with its built-in MTP head at depth 2. The same standard-v2 suite was used as for the other models: 400 MMLU-Pro items, 50 GSM8K items, and six coding tasks; temperature 0; 2,048-token output cap; four parallel requests. “At cap” counts identify truncated answers. The primary result below is specifically the thinking-off run.
Setup. AMD Radeon AI PRO R9700 (32 GB GDDR6); desktop AM5 platform, DDR5-6000, Windows 11; llama.cpp llama-server, Vulkan backend, build b11836; flash attention on; 131K context; ubatch 2048; four parallel slots sharing one unified KV cache. Tested October 2026.
| Sage-A1.5 run | MMLU-Pro | GSM8K | Coding | Decode | Suite wall time |
|---|---|---|---|---|---|
| Q4_K_M, thinking off, built-in MTP d2 | 86.75% (29 at cap) | 98% (1 at cap) | 6/6 | 195 tok/s | 59 min |
| Q4_K_M, thinking on, greedy | 55.25% (170 at cap) | 26% (38 at cap) | 0/6 | 198 tok/s | 189 min |
The near-identical decode rates make the distinction especially clear: throughput alone did not predict task performance. In the thinking-on run, the model spent tokens checking and rechecking its own work. One simple GSM8K prompt consumed the full 8,192-token available reasoning/output budget without producing an answer. The thinking-off rerun answered it correctly in roughly 230 tokens.
With thinking left on but using the model card's sampling parameters, it answered the same item correctly in about 3,400 tokens. That response still exceeded the suite's 2,048-token cap. The card's recommended sampling and our suite's greedy decoding serve different purposes. The suite holds temperature at zero for standardized comparisons; it does not represent the model card's recommended operating mode. In this case, the additional controlled observation suggests sampling reduced the loop for the simple item, but still used more tokens than disabling thinking. We did not rerun the full benchmark suite with the card's sampling settings, so there is no full-suite score for that configuration.
Recommended run settings
For benchmark reproduction, use the first row's configuration: Q4_K_M, MTP depth 2, 131K context, four-way parallelism, ubatch 2048, and thinking disabled. For ordinary model-card-faithful generation, start from temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 1.5, and repetition penalty 1.0.
Treat the thinking setting as a separate deliberate choice. If a short, bounded answer matters more than long internal exploration, turn thinking off; for multi-step reasoning or tool use, test thinking enabled with an adequate output budget and watch for loops. The card's native context is much larger than our tested 131K, but this benchmark does not establish quality at 256K.
Sage's built-in MTP also delivered a separate short-run speed benefit: at depth 2 we measured 207/174 tok/s on code/prose prompts, compared with 168 tok/s without drafting. These decode-sweep observations are separate from the multi-task suite figures above and do not establish the same quality result at a longer generated answer length. We found depth 2 preferable to deeper drafting in the short sweep.
Verdict
Sage-A1.5 is worth trying on this hardware when its mode is matched to the task. Thinking off gave our best observed Sage score profile and the fastest completed suite. Thinking on can spend an unexpectedly large budget in self-verification; temperature and sampling affect that behavior, but the card-recommended sample was not enough to fit the trivial answer under our 2,048-token cap.
Start with thinking off for concise work, then enable reasoning for tasks that need it and set explicit output limits. Do not compare these results without stating the mode.
Related reading: “MiniCPM-V 4.7 on the R9700: a fast multimodal MoE, with a DSpark caveat”; “Swift 1.5 versus stock Qwen3.8-27B on the R9700”; “Qwen3.8-27B on the R9700: MTP depth and quantization trade-offs”; “Qwen3.8-27B GSQ-RCO IQ3_S: 11.3 GiB and a useful quality floor”; and “Flint 40B-A2.5B on the R9700: fast decode, longer work”.
Disclaimer. These are measurements from one machine, one server build, and one test suite on the stated date. Results vary with hardware, software version, settings, and load. They are not a general ranking or a recommendation.