November 2026 · @migtissera · Continuity Research · LoRA adapter · Tess-4.5-27B
88.2% on SWE-Bench Pro v2 Hard · 9% fewer output tokens · 1M-token context · 594B parameters, 17B active · Apache 2.0
We're introducing Tess-4.5-594B-A17B, the Tess agentic-coding recipe on a frontier-scale base. It is Thinking Machines' Inkling 2 Small, a 594-billion-parameter Mixture-of-Experts model with 17 billion active parameters per token and a million-token window, post-trained on real engineering sessions to do what Tess does best: intellectually dense reasoning, warm and empathic tone, and hardcore engineering.
On SWE-Bench Pro v2 Hard it resolves more of the hardest public software-engineering tasks than the base model, and it does it with fewer tokens. On long-context reasoning it gives nothing back. It is a drop-in replacement for Inkling 2 Small: same checkpoint format, same serving recipes, same hardware.
| 88.2% | −9% | 1M | 594B · 17B |
|---|---|---|---|
| SWE-Bench Pro v2 Hard, against 86.3% for the base.¹ | Output tokens on those 51 tasks, against the base on the same harness.¹ | Tokens of context, the full window of the base. | Total and active parameters, Mixture-of-Experts. |
Three benchmarks from the Inkling 2 Small model card are public with open harnesses. We ran them the way the card describes, at reasoning_effort: "xhigh", and compare against the card's own numbers for the base.
| Tess-4.5-594B-A17B | Inkling 2 Small | |
|---|---|---|
| SWE-Bench Pro v2 Hard¹ | 88.2% | 86.3% |
| LongBench v2, 503 questions at 1M context² | 70.2% | 70.6% |
| AA-LCR v1.1³ | 85.0% | 83.3% |
The SWE-Bench Pro set is the one that separates frontier models: 51 tasks that at least two of five frontier model families failed under Scale's locked protocol. Tess resolves 45 of them. Across those 51 tasks it generated 4.07M output tokens where the base, run through the identical harness, generated 4.45M: 9% fewer, for more resolved tasks. Fewer tokens means answers sooner, less of the window spent on the model's own thinking, and longer sessions before compaction.
Run it at xhigh. Effort moves this model more than it moves the base: at high it scored 70.6% on the same 51 tasks.
Inkling 2 Small was available for fine-tuning on Tinker weeks before its weights were published. Tess-4.5-594B-A17B was trained there, on 46.9M tokens of anonymized real agentic coding sessions, and the LoRA was kept out of the experts on purpose: the released checkpoint stores its experts in NVFP4 and MXFP8, and a LoRA that touches them cannot be merged into them exactly.
When the weights arrived, the adapter was merged into the BF16 attention projections and output head, exactly, in FP32. Everything else in the checkpoint is byte-for-byte the base model: the quantized experts, the routers, the embeddings, the vision tower and the bundled DSpark draft. We then served the merged checkpoint and scored it token for token against the Tinker-served adapter on 897K held-out tokens: the two agree to within kernel noise.⁴ What you download is what we measured.
| Use it when | ||
|---|---|---|
| This repository | You want the model as a single checkpoint: the base model's format, about 394 GB, served by the official Inkling 2 Small vLLM and SGLang recipes without changes. | Merged, NVFP4/MXFP8 experts, BF16 attention |
| Tess-4.5-594B-A17B-LoRA | You already serve Inkling 2 Small and want to switch adapters at request time, or you want to inspect or re-merge the fine-tune yourself. 327 MB. | PEFT adapter, attention projections and head, rank 32 |
Hardware is the base model's: at least 500 GB of aggregated GPU memory, for example two B300, four H200, or four B200 at TP4. We ran the evaluations on four B200.
Give Tess its bearings. State the working directory in the system prompt and attach your real tool definitions. Tess reads the definitions and uses what is there; with no tools declared it answers in text.
You are Tess, an agentic coding assistant that operates in the user's terminal.
# Environment
Working directory: /path/to/your/project
tools field.reasoning_effort: "xhigh". Keep earlier-turn reasoning in the conversation in agent loops; the model was trained with it.vLLM, with the vllm/vllm-openai:inkling-2-small image until Inkling 2 Small support is in a release:
vllm serve migtissera/Tess-4.5-594B-A17B \
--tensor-parallel-size 4 \
--max-model-len 262144 \
--reasoning-parser inkling \
--enable-auto-tool-choice --tool-call-parser inkling \
--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"probabilistic"}'
SGLang, with the lmsysorg/sglang:dev-inkling-2-small image (eight H200 shown; see the SGLang cookbook for Blackwell recipes):
sglang serve --model-path migtissera/Tess-4.5-594B-A17B \
--tp 8 --ep 8 --attention-backend fa4 --dtype bfloat16 --page-size 128 \
--grammar-backend xgrammar --reasoning-parser inkling --tool-call-parser inkling2 \
--host 0.0.0.0 --port 30000
from openai import OpenAI
client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="migtissera/Tess-4.5-594B-A17B",
messages=[
{"role": "system", "content": "You are Tess, an agentic coding assistant.\n\n# Environment\n\nWorking directory: /home/me/project"},
{"role": "user", "content": "Why is the test suite failing?"},
],
tools=[{
"type": "function",
"function": {
"name": "Bash",
"description": "Executes a bash command and returns its output.",
"parameters": {"type": "object", "properties": {"command": {"type": "string"}}, "required": ["command"]},
},
}],
reasoning_effort="xhigh",
)
print(response.choices[0].message)
| Tensor group | In this repository |
|---|---|
Attention projections (wq_du, wk_dv, wv_dv, wo_ud, wr_du, gate_attn), all 48 layers |
Fine-tuned, BF16, exact merge |
| Unembedding head | Fine-tuned, BF16, exact merge |
| Routed experts (NVFP4 in layers 0–39, MXFP8 in layers 40–47), routers, attention-residual queries | Identical to the base |
| Embeddings, norms, short convolutions, vision tower, DSpark draft | Identical to the base |
Merging a LoRA into 4-bit expert weights would mean requantizing them, which erases most of a small update and perturbs the rest. Keeping the adapter on the BF16 tensors makes the merge exact (W ← W + (α/r)·BA, computed in FP32), keeps the checkpoint at the base model's size, and keeps every published serving recipe valid.
| Data | Tess-4.5 v2: anonymized real agentic coding sessions, Tess-4 samples and a self-distilled multi-turn conversational slice; 1,073 training rows of up to 64K tokens, 46.9M tokens, every row with the Tess system prompt, a working directory and the real tool definitions |
| Method | Supervised fine-tuning, LoRA rank 32 on the attention projections and the unembedding head (experts frozen), α = 32 |
| Schedule | One epoch, 135 steps of 8 rows, learning rate 1e-4 with linear decay, AdamW |
| Rendering | Thinking Machines' TML renderer at reasoning effort 0.9 (high), loss on every assistant message including its reasoning |
| Platform | Tinker |
| Held-out NLL | 0.683 (base) → 0.644 (Tess), token-weighted over 897K assistant tokens in 57 held-out sessions |
An agent that can run commands can also run the wrong one. Tess is built to ask before destructive actions and to report failures plainly, but no model is a guarantee. Run it with permission prompts, in a sandbox, or in a container. The weights are open; the responsibility for how they are deployed is yours.
Tess-4.5-594B-A17B is available under the Apache 2.0 license, built on Inkling 2 Small.
| This repository | Merged checkpoint in the base model's format: NVFP4/MXFP8 experts, BF16 attention; about 394 GB |
| Tess-4.5-594B-A17B-LoRA | The adapter alone, PEFT format |
| Base model | Inkling 2 Small: 594B total, 17B active, 48 layers, 384 routed experts (top-8), relative-bias attention, attention residuals, bundled DSpark speculator |
| Context | 1,048,576 tokens, inherited from the base |
| Reasoning | Effort-controlled thinking (reasoning_effort); run at xhigh |
| Vision | Inherited from the base; not part of the fine-tuning data |
| GGUF | Not yet: llama.cpp does not support the Inkling 2 architecture. A GGUF repository will follow when it does |
A note on the parameter count. Hugging Face's sidebar derives its count from tensor shapes, and the NVFP4 experts are stored two values per byte, so it under-reads them. The model has 594B parameters, 17B active per token, as the base does.
Limitations. Tess-4.5-594B-A17B is built for software engineering and agentic work in English. General chat, creative writing and image understanding come from the base model and were not a focus.
hard split of ScaleAI/SWE-bench_Pro) under Scale's locked protocol: mini-swe-agent 2.4.6 with Scale's tool-call config, agent sandboxes without network, one attempt per task, the agent's diff re-graded on a pristine image. Tess served by SGLang on four B200 (TP4/EP4, 1M context), temperature 1.0, top-p 0.95. The token figures are from the base run through the same harness on Tinker's endpoint, which scored 78.4% (40/51) at both high and xhigh, below its card number; the card's 86.3% is used above as the fairer reference for the base. Task by task against that run at xhigh, both models resolve 38, Tess alone 7, the base alone 2, neither 4. A 51-task set moves by a task or two on luck alone.Tess-4.5-594B-A17B stands on Inkling 2 Small. Thank you to Thinking Machines Lab for an exceptional open base model and for Tinker, which trained this adapter before the weights were public.
@misc{tess45_594b,
title = {Tess-4.5-594B-A17B},
author = {Migel Tissera},
year = {2026},
url = {https://huggingface.co/migtissera/Tess-4.5-594B-A17B}
}
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 GB