thinkingmachines/Inkling-2-Small (594B total / 17B active MoE, text and images in) with most of its refusals removed by abliteration. The files, format and numerics are the official release's: NVFP4 experts in layers 0-39, MXFP8 experts in layers 40-47, everything else BF16, and the same config, tokenizer and chat template. It runs anywhere the base model runs.
| base | this model | |
|---|---|---|
| KL from the base, first answer token | - | 0.249 |
| MMLU (14,042, 5-shot) | 79.3% | 80.3% |
This is the second version: a second abliteration pass over the first, which refuses less
(on 633 held-out prompts sampled 5 times each, 28.8% of replies refused or gave nothing
useful, against 31.7%; it also stops refusing most copyright requests). The first version
stays available at the v2-trial47 revision. Its full results follow; the benchmarks
below MMLU were not repeated on this version:
| base | first version | |
|---|---|---|
| KL from the base, first answer token | - | 0.161 |
| MMLU (14,042, 5-shot) | 79.3% | 80.0% |
| Global-MMLU-Lite (9,200, 23 languages) | 87.3% | 87.1% |
| GPQA Diamond (198, reasoning on) | 85.9% | 86.4% |
| LongBench v2 (503, 128k), see below | 54.1% | 48.5% |
| HumanEval (164) | 95.1% | 97.0% |
| MBPP (500) | 90.8% | 88.4% |
On the first version every difference but LongBench v2's is within noise; the next largest, MBPP's 2.4 points, is 23 problems lost against 11 gained.
LongBench v2 (first version): longer reasoning on long documents. With reasoning effort "none" and 2,048 tokens per reply, the first version runs out of tokens before answering more often than the base (178 of 503 against 131), which is where its 5.6 points go: on the questions both answered it is as accurate (185 right against 179). Given 8,192 tokens for those it reached an answer on 129 of 202, and on the 372 questions the base answered it gets 218 right against the base's 225. This version's long-context behaviour has not been measured; expect it to think longer than the base there as well until it is.
The same way as the base model; swap in this repo's id. See the base model's recipes for vLLM and SGLang. It needs the same hardware: the weights are 394 GB.
All numbers were measured by me, on both models with the same harness, prompts and greedy decoding (except the refusal comparison above, which samples at temperature 1.0). They compare the two models; they are not comparable to Thinking Machines' published scores, which use other harnesses and higher reasoning effort.
mmlu_generative, 5-shot. Global-MMLU-Lite, GPQA Diamond and LongBench v2:
zero-shot, generative, the answer letter read from the reply. HumanEval and MBPP:
zero-shot, the code taken from the reply's first code block and run against the tests.This model will help with requests the base model declines. You are responsible for how you use it and for complying with the law and with Thinking Machines' acceptable use policy. Anything the base model's card says about its risks and limitations applies here, and its safety evaluations do not.
model-00022-of-00024.safetensors 20 GBmodel-00002-of-00024.safetensors 19 GBmodel-00004-of-00024.safetensors 19 GBmodel-00012-of-00024.safetensors 19 GBmodel-00010-of-00024.safetensors 19 GBmodel-00008-of-00024.safetensors 19 GBmodel-00006-of-00024.safetensors 19 GBmodel-00014-of-00024.safetensors 19 GBmodel-00023-of-00024.safetensors 19 GBmodel-00007-of-00024.safetensors 17 GBmodel-00009-of-00024.safetensors 17 GBmodel-00013-of-00024.safetensors 17 GBmodel-00005-of-00024.safetensors 17 GBmodel-00003-of-00024.safetensors 17 GBmodel-00011-of-00024.safetensors 17 GBmodel-00001-of-00024.safetensors 16 GBmodel-00015-of-00024.safetensors 15 GBmodel-00021-of-00024.safetensors 13 GBmodel-00016-of-00024.safetensors 13 GBmodel-00017-of-00024.safetensors 13 GBmodel-00018-of-00024.safetensors 13 GBmodel-00019-of-00024.safetensors 13 GBmodel-00020-of-00024.safetensors 13 GBmodel-00024-of-00024.safetensors 12 GBthinkingmachines/Inkling-2-Small