November 2026 · Merged model · Continuity Research
This is the LoRA that turns Thinking Machines' Inkling 2 Small into Tess-4.5-594B-A17B: 327 MB in PEFT format, trained on Tinker on 46.9M tokens of real agentic engineering sessions before the base weights were public. Merged exactly into the base checkpoint it scores 88.2% on SWE-Bench Pro v2 Hard against the base's 86.3%, with 9% fewer output tokens, and stays level with the base on long-context reasoning. The numbers, the methodology and the footnotes are on the merged model's card.
| Base model | thinkingmachines/Inkling-2-Small |
| Rank / α | 32 / 32 (scaling 1.0) |
| Modules | The six attention projections of all 48 layers (wq_du, wk_dv, wv_dv, wo_ud, wr_du, gate_attn) and the unembedding head (lm_head); the routed experts are not adapted |
| Parameters | 81.7M (577 tensors, FP32, 327 MB) |
| Trained with | Tinker: one epoch of the Tess-4.5 v2 data, learning rate 1e-4 with linear decay, batch 8, effort 0.9 rendering |
The released checkpoint stores its routed experts in NVFP4 and MXFP8. A LoRA that touches them cannot be merged into them exactly, and vLLM cannot serve expert adapters on top of quantized experts either. An attention-and-head adapter does both: it merges exactly into the BF16 tensors (that merge is the merged repository, verified token for token against the Tinker-served adapter), and it is the kind of adapter the base model's serving recipes accept.
The merged model is the tested path. Download Tess-4.5-594B-A17B unless you need one of the two things below.
Switch adapters at request time. vLLM's Inkling 2 Small recipe supports LoRA on the attention projections of the quantized checkpoint:
vllm serve thinkingmachines/Inkling-2-Small \
--tensor-parallel-size 4 --max-model-len 262144 \
--enable-lora --lora-modules tess=migtissera/Tess-4.5-594B-A17B-LoRA \
--reasoning-parser inkling --enable-auto-tool-choice --tool-call-parser inkling
Request the adapter by name (model="tess"), with reasoning_effort="xhigh". We have not load-tested adapter serving ourselves.
Merge it yourself. For every lora_A/lora_B pair the target is model.llm.layers.<n>.attn.<proj>.weight (language_model.lm_head → model.llm.unembed.weight), all BF16 in the base checkpoint: W ← bf16(fp32(W) + (α/r)·B·A). Everything else is copied unchanged.
Tess expects a system prompt that states the working directory, the real tool definitions in the tools field, and reasoning_effort: "xhigh". See the merged model's card for the full quickstart.
Built on Inkling 2 Small by Thinking Machines Lab, and trained with Tinker.
@misc{tess45_594b,
title = {Tess-4.5-594B-A17B},
author = {Migel Tissera},
year = {2026},
url = {https://huggingface.co/migtissera/Tess-4.5-594B-A17B}
}
thinkingmachines/Inkling-2-Small