# Unbind Flash V1 — benchmark comparisons

Unbind Flash V1 launch graphics. The highlighted “Unbind Flash V1 · base model” row represents the unmodified DeepSeek-V4.1-Flash-1017 reference checkpoint, using scores reported by DeepSeek. These are base-model comparisons, not evaluations of Unbind’s endpoint, abliterated hosted checkpoint or older public GGUF.

Each chart compares models within one benchmark on a 0–100 axis. Evaluation setups differ across benchmarks; do not average these scores into an overall rating.

[Source table and evaluation notes](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017).

## Animated overview

![Eight benchmark comparisons](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/benchmark-overview.gif)

## DeepSWE v1.1

![DeepSWE v1.1 comparison](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/01-deepswe.png)

Mini-SWE-V2 harness. Publisher-reported benchmark scores.

## Terminal-Bench 4.0

![Terminal-Bench 4.0 comparison](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/02-terminal.png)

Claude Code harness. Publisher-reported benchmark scores.

## HLE without tools

![HLE without tools comparison](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/03-hle.png)

Full HLE figures only. GLM-5.3 omitted: the source gives a text-only subset score.

## ProgramBench

![ProgramBench comparison](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/04-program.png)

Publisher-reported benchmark scores. Additional per-row settings are not specified.

## NL2Repo-Bench

![NL2Repo-Bench comparison](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/05-nl2repo.png)

Publisher-reported benchmark scores. Additional per-row settings are not specified.

## Automation-Bench v1.0.6

![Automation-Bench v1.0.6 comparison](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/06-automation.png)

Publisher-reported benchmark scores. Additional per-row settings are not specified.

## Agents’ Last Exam

![Agents’ Last Exam comparison](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/07-agents.png)

Publisher-reported benchmark scores. Additional per-row settings are not specified.

## Apex Bench v1.1

![Apex Bench v1.1 comparison](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/08-apex.png)

Publisher-reported benchmark scores. Additional per-row settings are not specified.

## Data and reuse

[CSV](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/source-data.csv) · [JSON](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/benchmarks/source-data.json). The accompanying SVG files provide editable vector exports.

All plotted values reproduce the source table. The highlighted display label is branded as “Unbind Flash V1 · base model”; its exact source model is retained in the footer and data files. Other model labels reproduce the source table. GLM-5.3 is omitted from HLE because only its text-only subset score is supplied. These are reported benchmark comparisons, not newly run evaluations.