The Neurion model family
Keep the work moving.
Explain a function, sketch an approach, refine a draft, or make a focused code change. Neurion Flash is the model to choose for the interactive back-and-forth that moves everyday work forward.
PreviewHydraLabs-reported results
How we evaluated
HydraLabs reports Neurion Flash results across coding, tool use, scientific reasoning and security tasks. The reported setup identifies Ultra, with MelissaCode 0.1.0 for applicable tool-enabled tasks and mini-SWE-agent for DeepSWE.
- Reported configuration
- Ultramanaged-agent workflow
- Context window
- 1Mtokens
- Maximum output
- 384Ktokens
Model capacity is separate from an evaluation’s time, token and tool budget. The scores do not establish a shared budget across models.
Neurion Flash scores are HydraLabs-reported internal results. Run-level reports recording task budgets, attempts and costs have not been published. Competitor results retain their attributed benchmark versions and configurations in Notes & sources; differing task sets, tools or settings can limit direct comparison.
How to read the resultsView the benchmark reportCoding Bench
DeepSWE v1.1
Notes & sources
Neurion Flash scores are HydraLabs-reported internal results. The reported setup identifies Ultra, MelissaCode 0.1.0 for applicable tool-enabled tasks and mini-SWE-agent for DeepSWE. Run-level reports recording task budgets, attempts and costs have not been published. Model capacity of 1M context tokens and 384K output tokens is separate from benchmark budgets. Competitor scores retain their attributed configurations; differing task sets, tools and settings can limit direct comparison. A dash means no matching result is available, never zero.
Qwen results identify Flash-Next, the basis of the Flash API. MiMo publisher results identify Flash-RL; independent evaluations identify the MiMo-V2.6-Flash API. GLM results are for GLM 5.3 Flash and DeepSeek results for the 0731 Flash release. Evaluated artifacts and configurations are recorded with each source.
HydraLabs-reported result · Effort: Ultra · mini-SWE-agent
HydraLabs-reported internal result
Publisher-reported result · Qwen/Qwen3.8-Flash-Next · mini-swe-agent (best of Claude Code and mini-swe-agent) · T=1, top_p=.95, 256K context. Publisher reports the better of two harnesses.
Publisher sourcePublisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Publisher table; run configuration not specified.
Publisher sourcePublisher-reported result · zai-org/GLM-5.3-Flash · Effort: Max · mini-swe-agent · T=.95, top_p=1, 400K context, six-hour timeout.
Publisher sourcePublisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · Effort: Max · DeepSeek Harness minimal · T=1, top_p=.95, Max. Exact card calls this DeepSWE without a revision; Qwen's separate comparison labels the same 54.4 baseline DeepSWE 1.1.
Publisher sourcePublisher-reported result · Effort: Max · OpenAI internal research harness · DeepSWE v1.1; publisher result at Max. Separate mini-SWE-agent evaluations are not substituted.
Publisher sourceHydraLabs-provided comparison · huggingface.co
HydraLabs-provided comparison · huggingface.co
HydraLabs-provided comparison · huggingface.co
Agentic Bench
Toolathlon Verified
Notes & sources
Neurion Flash scores are HydraLabs-reported internal results. The reported setup identifies Ultra, MelissaCode 0.1.0 for applicable tool-enabled tasks and mini-SWE-agent for DeepSWE. Run-level reports recording task budgets, attempts and costs have not been published. Model capacity of 1M context tokens and 384K output tokens is separate from benchmark budgets. Competitor scores retain their attributed configurations; differing task sets, tools and settings can limit direct comparison. A dash means no matching result is available, never zero.
Qwen results identify Flash-Next, the basis of the Flash API. MiMo publisher results identify Flash-RL; independent evaluations identify the MiMo-V2.6-Flash API. GLM results are for GLM 5.3 Flash and DeepSeek results for the 0731 Flash release. Evaluated artifacts and configurations are recorded with each source.
No result supplied for: GPT-6 Luna.
HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
HydraLabs-reported internal result
Publisher-reported result · Qwen/Qwen3.8-Flash-Next · Pass@1.
Publisher sourcePublisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Publisher table; run configuration not specified.
Publisher sourcePublisher-reported result · zai-org/GLM-5.3-Flash · Effort: Max · Official evaluation service · Pass@1 averaged over three independent runs.
Publisher sourcePublisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · Publisher table; run configuration not specified.
Publisher sourceHydraLabs-provided comparison · huggingface.co
HydraLabs-provided comparison · huggingface.co
Reasoning Bench
GPQA Diamond
Notes & sources
Neurion Flash scores are HydraLabs-reported internal results. The reported setup identifies Ultra, MelissaCode 0.1.0 for applicable tool-enabled tasks and mini-SWE-agent for DeepSWE. Run-level reports recording task budgets, attempts and costs have not been published. Model capacity of 1M context tokens and 384K output tokens is separate from benchmark budgets. Competitor scores retain their attributed configurations; differing task sets, tools and settings can limit direct comparison. A dash means no matching result is available, never zero.
Qwen results identify Flash-Next, the basis of the Flash API. MiMo publisher results identify Flash-RL; independent evaluations identify the MiMo-V2.6-Flash API. GLM results are for GLM 5.3 Flash and DeepSeek results for the 0731 Flash release. Evaluated artifacts and configurations are recorded with each source.
Luna 90.47, GLM Flash 90.15 and DeepSeek Flash 0731 91.04 use exact Epoch model-version rows at Max. Qwen 91.7 is publisher-reported Flash-Next. MiMo has no matching verified result.
No result supplied for: MiMo V2.6 Flash.
HydraLabs-reported result · Effort: Ultra
HydraLabs-reported internal result
Publisher-reported result · Qwen/Qwen3.8-Flash-Next · Publisher result for Flash-Next, the basis of the Flash API. Evaluation harness, reasoning effort and exact API snapshot were not specified.
Publisher sourceIndependent evaluation by Epoch AI · glm-5.3-flash_max · Effort: Max · Evaluation date: 2026-08-26 · Exact model-version row from Epoch gpqa_diamond.csv; mean_score converted to a percentage and rounded to two decimals.
Epoch AI · Independent evaluationIndependent evaluation by Epoch AI · deepseek-v4-flash-0731_max · Effort: Max · Evaluation date: 2026-08-02 · Exact model-version row from Epoch gpqa_diamond.csv; mean_score converted to a percentage and rounded to two decimals.
Epoch AI · Independent evaluationIndependent evaluation by Epoch AI · gpt-6-luna_max · Effort: Max · Evaluation date: 2026-09-22 · Exact model-version row from Epoch gpqa_diamond.csv; mean_score converted to a percentage and rounded to two decimals.
Epoch AI · Independent evaluationCyber Bench
CyberGym
Notes & sources
Neurion Flash scores are HydraLabs-reported internal results. The reported setup identifies Ultra, MelissaCode 0.1.0 for applicable tool-enabled tasks and mini-SWE-agent for DeepSWE. Run-level reports recording task budgets, attempts and costs have not been published. Model capacity of 1M context tokens and 384K output tokens is separate from benchmark budgets. Competitor scores retain their attributed configurations; differing task sets, tools and settings can limit direct comparison. A dash means no matching result is available, never zero.
Qwen results identify Flash-Next, the basis of the Flash API. MiMo publisher results identify Flash-RL; independent evaluations identify the MiMo-V2.6-Flash API. GLM results are for GLM 5.3 Flash and DeepSeek results for the 0731 Flash release. Evaluated artifacts and configurations are recorded with each source.
No result supplied for: Qwen 3.8 Flash, GLM 5.3 Flash, GPT-6 Luna.
HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
HydraLabs-reported internal result
Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Publisher table; run configuration not specified.
Publisher sourcePublisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · Publisher table; run configuration not specified.
Publisher sourceMath Bench
FrontierMath Tier 4 (v2)
Notes & sources
Neurion Flash scores are HydraLabs-reported internal results. The reported setup identifies Ultra, MelissaCode 0.1.0 for applicable tool-enabled tasks and mini-SWE-agent for DeepSWE. Run-level reports recording task budgets, attempts and costs have not been published. Model capacity of 1M context tokens and 384K output tokens is separate from benchmark budgets. Competitor scores retain their attributed configurations; differing task sets, tools and settings can limit direct comparison. A dash means no matching result is available, never zero.
Qwen results identify Flash-Next, the basis of the Flash API. MiMo publisher results identify Flash-RL; independent evaluations identify the MiMo-V2.6-Flash API. GLM results are for GLM 5.3 Flash and DeepSeek results for the 0731 Flash release. Evaluated artifacts and configurations are recorded with each source.
Exact Epoch model-version rows at Max. Raw mean_score values are converted to percentages and rounded to two decimals.
No result supplied for: Neurion Flash, Qwen 3.8 Flash, MiMo V2.6 Flash.
Independent evaluation by Epoch AI · glm-5.3-flash_max · Effort: Max · Evaluation date: 2026-08-27 · FrontierMath Tier 4 v2 at Max effort. The exact Epoch model-version row supplies the run date, raw mean and standard error. Percentage equals the raw mean multiplied by 100, rounded to two decimal places.
Epoch AI · Independent evaluationIndependent evaluation by Epoch AI · deepseek-v4-flash-0731_max · Effort: Max · Evaluation date: 2026-08-02 · FrontierMath Tier 4 v2 at Max effort. The exact Epoch model-version row supplies the run date, raw mean and standard error. Percentage equals the raw mean multiplied by 100, rounded to two decimal places.
Epoch AI · Independent evaluationIndependent evaluation by Epoch AI · gpt-6-luna_max · Effort: Max · Evaluation date: 2026-09-22 · FrontierMath Tier 4 v2 at Max effort. The exact Epoch model-version row supplies the run date, raw mean and standard error. Percentage equals the raw mean multiplied by 100, rounded to two decimal places.
Epoch AI · Independent evaluationAutonomous behavior and task boundaries
Define the permitted targets and actions before working with tools or agents. Keep access limited to the task, review the activity and retain a way to pause the run.
Flash-specific evaluation observations will be added with the results.
Explore, build, and play YABAN, a single-player voxel sandbox.
Games Performance
YABAN was developed by Neurion Flash in a total of 1 hour and 40 minutes. During development, the model identified and corrected errors on 12 occasions entirely on its own. There was no team intervention at any stage.
The initial prompt consisted of just three lines. The model used web research during development. Stating that it wanted to present its work honestly, it independently chose not to use ready-made game assets.
After development was completed, no corrections were applied by the team. There are no known issues in the current version.
3D Performance
Rotate each model, zoom in, and inspect the details.
Explore the Aurora TFX-01 cutaway. Rotate the model and zoom in to see its internal layout.
Put Flash to work.
Choose Neurion Flash in the model selector and start in Chat or Code. Model choice and task workflow are separate controls; Ultra coordinates managed agents using the selected model.
Neurion Flash is included in the Core, Security, Pro and Max subscription plans, within the selected workspace’s shared usage limits.
The full benchmark report.
Coding, agentic, reasoning, cyber and math results.
HydraLabs-reported Flash results alongside publisher and independent evaluations. Unavailable results remain —.
| Benchmark | Neurion Flash | Qwen 3.8 Flash | MiMo V2.6 Flash | GLM 5.3 Flash | DeepSeek V4 Flash 0731 | GPT-6 Luna |
|---|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 24.7% — Ultra | 25% | 28.8% | 35.8% — None | 12% — Max | 16.4% — Max |
| Terminal-Bench 2.1 | 85.2% — Ultra | — | 87.6% | 84.3% — Max | 82.7% — Max | — |
| NL2Repo | 58.2% — Ultra | 48.1% | — | 56.3% — Max | 54.2% — Max | — |
| DeepSWE v1.1 | 62.3% — Ultra | 58.7% | 67.9% | 63.4% — Max | 54.4% — Max | 66.6% — Max |
| SciCode | — | 51% | 51% | 51.6% | 50% — Max | 54.6% — Max |
| GDPval-AA v2.1 | — | 1629 Elo | 1608 Elo | 1644 Elo | 1434 Elo — Max | 1432 Elo — Max |
| AA-Briefcase v1.1 | — | 1581 Elo | 1493 Elo | 1454 Elo | 1259 Elo — Max | 1336 Elo — Max |
| Toolathlon Verified | 73.7% — Ultra | 73.5% | 73.6% | 78.4% — Max | 70.3% | — |
| ProgramBench · Almost Resolved | — | — | 17.5% | — | — | — |
| Agents’ Last Exam | 39.2% — Ultra | 24.3% | 27.6% | 26.3% — Max | 25.2% | — |
| Agents’ Last Exam V1 · 55 sub-industries | — | — | — | — | — | 50.9% — Max |
| AutomationBench | 46.7% — Ultra | — | 52.3% | 48.8% — Max | 25.1% | 20.7% — Max |
| Humanity’s Last Exam · Text-only, without tools | — | 38% | 35% | 39.9% | 39% — Max | 38.5% — Max |
| Humanity’s Last Exam · Full set, with tools | — | — | — | 55.3% — Max | — | — |
| GPQA Diamond | 88.7% — Ultra | 91.7% | — | 90.15% — Max | 91.04% — Max | 90.47% — Max |
| Terminal-Bench Science 0.1 | — | — | 4.29% | 2.86% — Max | 4.29% — High | — |
| CyberGym | 81.3% — Ultra | — | 95.1% | — | 76.7% | — |
| SEC-Bench Pro | 43.5% — Ultra | — | 47.5% | — | — | 34.2% |
| ExploitGym | 12.3% — Ultra | — | 6% | — | — | 11.6% |
| ExploitBench | 28.4% — Ultra | — | 25.3% | — | — | — |
| FrontierMath Tier 4 (v2) | — | — | — | 17.07% — Max | 24.39% — Max | 56.1% — Max |
Evaluation notes & sources
Neurion Flash scores are HydraLabs-reported internal results. The reported setup identifies Ultra, MelissaCode 0.1.0 for applicable tool-enabled tasks and mini-SWE-agent for DeepSWE. Run-level reports recording task budgets, attempts and costs have not been published. Model capacity of 1M context tokens and 384K output tokens is separate from benchmark budgets. Competitor scores retain their attributed configurations; differing task sets, tools and settings can limit direct comparison. A dash means no matching result is available, never zero.
Qwen results identify Flash-Next, the basis of the Flash API. MiMo publisher results identify Flash-RL; independent evaluations identify the MiMo-V2.6-Flash API. GLM results are for GLM 5.3 Flash and DeepSeek results for the 0731 Flash release. Evaluated artifacts and configurations are recorded with each source.
- Terminal-Bench 4.0
Snorkel lists Luna at Max in Codex and GLM Flash with effort “None” in Claude Code. DeepSeek uses Artificial Analysis’s Max / mini-swe-agent run. Model release dates on the Snorkel board are not evaluation dates.
Return to resultNeurion Flash · 24.7%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
Qwen 3.8 Flash · 25%Independent evaluation by Artificial Analysis · Qwen/Qwen3.8-Flash-Next · mini-swe-agent · Terminal-Bench 4.0; 66 tasks; pass@1 averaged over three repeats. mini-swe-agent uses at most 500 steps, upstream task timeouts and the complete transcript without compaction.
Artificial Analysis · Independent evaluationMiMo V2.6 Flash · 28.8%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Flash-RL publisher result on Terminal-Bench 4.0. Harness, reasoning effort and evaluation date are unspecified.
Publisher sourceGLM 5.3 Flash · 35.8%Independent evaluation by Snorkel AI · Effort: None · Claude Code · Terminal-Bench4.0;66 tasks;35.8% ±3.5; effort is None as explicitly listed by this evaluator, not the vendor Max setting.
Snorkel AI · Independent evaluationDeepSeek V4 Flash 0731 · 12%Independent evaluation by Artificial Analysis · deepseek-ai/DeepSeek-V4-Flash-0731 · Effort: Max · mini-swe-agent · Exact 0731 Flash release. Full 66-task set, pass@1 across three repeats, at most 500 agent steps, upstream timeouts and full transcripts without compaction.
Artificial Analysis · Independent evaluationGPT-6 Luna · 16.4%Independent evaluation by Snorkel AI · Effort: Max · Codex · Terminal-Bench4.0;66 tasks;16.4% ±2.7. Date is the model release date in the board, not an inferred run date.
Snorkel AI · Independent evaluation - Terminal-Bench 2.1
Terminal tasks on Terminal-Bench 2.1. These scores are not interchangeable with version 4.0.
Return to resultNeurion Flash · 85.2%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
MiMo V2.6 Flash · 87.6%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Publisher table; run configuration not specified.
Publisher sourceGLM 5.3 Flash · 84.3%Publisher-reported result · zai-org/GLM-5.3-Flash · Effort: Max · Claude Code 2.1.207 · T=1, top_p=1, max_new_tokens=65536, six-hour timeout.
Publisher sourceDeepSeek V4 Flash 0731 · 82.7%Publisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · Effort: Max · DeepSeek Harness minimal · T=1, top_p=.95. This is publisher 2.1, not AA's independent 2.1 or 4.0 run.
Publisher source - NL2Repo
Building a working repository from natural-language requirements.
Return to resultNeurion Flash · 58.2%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
Qwen 3.8 Flash · 48.1%Publisher-reported result · Qwen/Qwen3.8-Flash-Next · Claude Code · Repository-download commands disabled by evaluator.
Publisher sourceGLM 5.3 Flash · 56.3%Publisher-reported result · zai-org/GLM-5.3-Flash · Effort: Max · T=1, top_p=1, max_new_tokens=64K, 1M context; behavioral validation applied.
Publisher sourceDeepSeek V4 Flash 0731 · 54.2%Publisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · Effort: Max · DeepSeek Harness minimal · Publisher table; run configuration not specified.
Publisher source - DeepSWE v1.1
Repository-level software engineering on DeepSWE v1.1.
Return to resultNeurion Flash · 62.3%HydraLabs-reported result · Effort: Ultra · mini-SWE-agent
Qwen 3.8 Flash · 58.7%Publisher-reported result · Qwen/Qwen3.8-Flash-Next · mini-swe-agent (best of Claude Code and mini-swe-agent) · T=1, top_p=.95, 256K context. Publisher reports the better of two harnesses.
Publisher sourceMiMo V2.6 Flash · 67.9%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Publisher table; run configuration not specified.
Publisher sourceGLM 5.3 Flash · 63.4%Publisher-reported result · zai-org/GLM-5.3-Flash · Effort: Max · mini-swe-agent · T=.95, top_p=1, 400K context, six-hour timeout.
Publisher sourceDeepSeek V4 Flash 0731 · 54.4%Publisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · Effort: Max · DeepSeek Harness minimal · T=1, top_p=.95, Max. Exact card calls this DeepSWE without a revision; Qwen's separate comparison labels the same 54.4 baseline DeepSWE 1.1.
Publisher sourceGPT-6 Luna · 66.6%Publisher-reported result · Effort: Max · OpenAI internal research harness · DeepSWE v1.1; publisher result at Max. Separate mini-SWE-agent evaluations are not substituted.
Publisher source - SciCode
Artificial Analysis currently marks SciCode as under review. Each result retains the precision published by its selected source.
Return to resultQwen 3.8 Flash · 51%Independent evaluation by Artificial Analysis · Qwen/Qwen3.8-Flash-Next · SciCode v1.0.1; 288 test subproblems with scientist-provided background; subproblem pass@1 over three repeats and a 300-second isolated execution timeout.
Artificial Analysis · Independent evaluationMiMo V2.6 Flash · 51%Independent evaluation by Artificial Analysis · mimo-v2.6-flash · SciCode v1.0.1; 288 test subproblems with scientist-provided background; subproblem pass@1 over three repeats and a 300-second isolated execution timeout.
Artificial Analysis · Independent evaluationGLM 5.3 Flash · 51.6%Independent evaluation by Artificial Analysis · glm-5.3-flash · SciCode v1.0.1; 288 test subproblems with scientist-provided background; subproblem pass@1 over three repeats and a 300-second isolated execution timeout.
Artificial Analysis · Independent evaluationDeepSeek V4 Flash 0731 · 50%Independent evaluation by Artificial Analysis · deepseek-v4-flash-0731 · Effort: Max · SciCode v1.0.1; 288 test subproblems with scientist-provided background; subproblem pass@1 over three repeats and a 300-second isolated execution timeout.
Artificial Analysis · Independent evaluationGPT-6 Luna · 54.6%Independent evaluation by Artificial Analysis · gpt-6-luna · Effort: Max · SciCode v1.0.1; 288 test subproblems with scientist-provided background; subproblem pass@1 over three repeats and a 300-second isolated execution timeout.
Artificial Analysis · Independent evaluation - GDPval-AA v2.1
Elo values retain their dated research snapshots; live Elo rankings can change. These are Elo scores, not percentages.
Return to resultQwen 3.8 Flash · 1629 EloIndependent evaluation by Artificial Analysis · Qwen/Qwen3.8-Flash-Next · GDPval-AA v2.1; 220 tasks and one run. Elo is calculated from judge-panel pairwise comparisons. This is an Elo rating, not a percentage.
Artificial Analysis · Independent evaluationMiMo V2.6 Flash · 1608 EloIndependent evaluation by Artificial Analysis · mimo-v2.6-flash · GDPval-AA v2.1; 220 tasks and one run. Elo is calculated from judge-panel pairwise comparisons. This is an Elo rating, not a percentage.
Artificial Analysis · Independent evaluationGLM 5.3 Flash · 1644 EloIndependent evaluation by Artificial Analysis · glm-5.3-flash · GDPval-AA v2.1; 220 tasks and one run. Elo is calculated from judge-panel pairwise comparisons. This is an Elo rating, not a percentage.
Artificial Analysis · Independent evaluationDeepSeek V4 Flash 0731 · 1434 EloIndependent evaluation by Artificial Analysis · deepseek-v4-flash-0731 · Effort: Max · GDPval-AA v2.1; 220 tasks and one run. Elo is calculated from judge-panel pairwise comparisons. This is an Elo rating, not a percentage.
Artificial Analysis · Independent evaluationGPT-6 Luna · 1432 EloIndependent evaluation by Artificial Analysis · gpt-6-luna · Effort: Max · GDPval-AA v2.1; 220 tasks and one run. Elo is calculated from judge-panel pairwise comparisons. This is an Elo rating, not a percentage.
Artificial Analysis · Independent evaluation - AA-Briefcase v1.1
Elo values retain their dated research snapshots; live Elo rankings can change. These are Elo scores, not percentages.
Return to resultQwen 3.8 Flash · 1581 EloIndependent evaluation by Artificial Analysis · Qwen/Qwen3.8-Flash-Next · AA-Briefcase v1.1; 91 tasks across four scenarios and one run. Combined Elo measures rubric success, analytical quality and presentation quality.
Artificial Analysis · Independent evaluationMiMo V2.6 Flash · 1493 EloIndependent evaluation by Artificial Analysis · mimo-v2.6-flash · AA-Briefcase v1.1; 91 tasks across four scenarios and one run. Combined Elo measures rubric success, analytical quality and presentation quality.
Artificial Analysis · Independent evaluationGLM 5.3 Flash · 1454 EloIndependent evaluation by Artificial Analysis · glm-5.3-flash · AA-Briefcase v1.1; 91 tasks across four scenarios and one run. Combined Elo measures rubric success, analytical quality and presentation quality.
Artificial Analysis · Independent evaluationDeepSeek V4 Flash 0731 · 1259 EloIndependent evaluation by Artificial Analysis · deepseek-v4-flash-0731 · Effort: Max · AA-Briefcase v1.1; 91 tasks across four scenarios and one run. Combined Elo measures rubric success, analytical quality and presentation quality.
Artificial Analysis · Independent evaluationGPT-6 Luna · 1336 EloIndependent evaluation by Artificial Analysis · gpt-6-luna · Effort: Max · AA-Briefcase v1.1; 91 tasks across four scenarios and one run. Combined Elo measures rubric success, analytical quality and presentation quality.
Artificial Analysis · Independent evaluation - Toolathlon Verified
Multi-tool workflow completion on Toolathlon Verified.
Return to resultNeurion Flash · 73.7%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
Qwen 3.8 Flash · 73.5%Publisher-reported result · Qwen/Qwen3.8-Flash-Next · Pass@1.
Publisher sourceMiMo V2.6 Flash · 73.6%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Publisher table; run configuration not specified.
Publisher sourceGLM 5.3 Flash · 78.4%Publisher-reported result · zai-org/GLM-5.3-Flash · Effort: Max · Official evaluation service · Pass@1 averaged over three independent runs.
Publisher sourceDeepSeek V4 Flash 0731 · 70.3%Publisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · Publisher table; run configuration not specified.
Publisher source - ProgramBench
Vals Almost Resolved requires at least 95% of hidden tests to pass. The fully resolved and raw test-pass metrics are separate.
Return to resultMiMo V2.6 Flash · 17.5%Independent evaluation by Vals AI · xiaomi/mimo-v2.6-flash · Mini-SWE-agent · ProgramBench Almost Resolved: at least 95% of hidden tests pass. 200 public tasks; mini-SWE-agent, offline cleanroom, 1,000 steps, six-hour wall-clock limit and 180-second action timeout. Temperature 1, top_p 0.95 and 128,000 maximum output tokens.
Vals AI · Independent evaluation - Agents’ Last Exam
¹ Qwen 24.3 is Pass@1, distinct from its separate 51.2 Score. The supplied task set is unspecified; this comparison is not relabeled ALE-CLI.
Return to resultNeurion Flash · 39.2%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
Qwen 3.8 Flash · 24.3%Publisher-reported result · Qwen/Qwen3.8-Flash-Next · The publisher comparison does not establish a common task set across models. 24.3 is Pass@1, separate from the publisher’s 51.2 Score.
Publisher sourceMiMo V2.6 Flash · 27.6%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · The publisher comparison does not establish a common task set across models.
Publisher sourceGLM 5.3 Flash · 26.3%Publisher-reported result · zai-org/GLM-5.3-Flash · Effort: Max · The publisher comparison does not establish a common task set across models. Its model card separately identifies 26.3 as an ALE CLI result.
Publisher sourceDeepSeek V4 Flash 0731 · 25.2%Publisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · The publisher comparison does not establish a common task set across models.
Publisher source - Agents’ Last Exam V1 · 55 sub-industries
Luna V1 scores are not labeled ALE-CLI or Pass@1. No matching Neurion Flash V1 task-set result was supplied.
Return to resultGPT-6 Luna · 50.9%Publisher-reported result · gpt-6-luna · Effort: Max · OpenAI release evaluation on Agents’ Last Exam V1 across 55 sub-industries. The metric is not identified as CLI or Pass@1. Harness, sampling, attempts and evaluation date are unspecified.
Publisher source - AutomationBench
MiMo and GLM identify public v1.0.6; Luna’s official release chart names v1.0.6. DeepSeek says Public without a revision. The Flash benchmark revision was not supplied.
Return to resultNeurion Flash · 46.7%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
MiMo V2.6 Flash · 52.3%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Publisher explicitly labels v1.0.6.
Publisher sourceGLM 5.3 Flash · 48.8%Publisher-reported result · zai-org/GLM-5.3-Flash · Effort: Max · Publisher explicitly labels v1.0.6 with null-type handling fix.
Publisher sourceDeepSeek V4 Flash 0731 · 25.1%Publisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · Publisher reports AutomationBench Public without a revision number. This is distinct from AutomationBench-AA.
Publisher sourceGPT-6 Luna · 20.7%Publisher-reported result · gpt-6-luna · Effort: Max · OpenAI release evaluation on AutomationBench 1.0.6 with 47 business tools. Harness, sampling, attempts and evaluation date are unspecified. Research results may differ from production behavior.
Publisher source - HLE · Text-only
Text-only May 2025 revision, 2,158 questions, pass@1, without tools. Full-set and tools-enabled evaluations are shown separately.
Return to resultQwen 3.8 Flash · 38%Independent evaluation by Artificial Analysis · Qwen/Qwen3.8-Flash-Next · Humanity's Last Exam; May 2025 revision, 2,158 text-only questions without tools. pass@1 from one repeat, graded with the GPT-5.6 Luna Medium equality checker.
Artificial Analysis · Independent evaluationMiMo V2.6 Flash · 35%Independent evaluation by Artificial Analysis · mimo-v2.6-flash · Humanity's Last Exam; May 2025 revision, 2,158 text-only questions without tools. pass@1 from one repeat, graded with the GPT-5.6 Luna Medium equality checker.
Artificial Analysis · Independent evaluationGLM 5.3 Flash · 39.9%Independent evaluation by Artificial Analysis · glm-5.3-flash · Humanity's Last Exam; May 2025 revision, 2,158 text-only questions without tools. pass@1 from one repeat, graded with the GPT-5.6 Luna Medium equality checker.
Artificial Analysis · Independent evaluationDeepSeek V4 Flash 0731 · 39%Independent evaluation by Artificial Analysis · deepseek-v4-flash-0731 · Effort: Max · Humanity's Last Exam; May 2025 revision, 2,158 text-only questions without tools. pass@1 from one repeat, graded with the GPT-5.6 Luna Medium equality checker.
Artificial Analysis · Independent evaluationGPT-6 Luna · 38.5%Independent evaluation by Artificial Analysis · gpt-6-luna · Effort: Max · Humanity's Last Exam; May 2025 revision, 2,158 text-only questions without tools. pass@1 from one repeat, graded with the GPT-5.6 Luna Medium equality checker.
Artificial Analysis · Independent evaluation - HLE · With tools
Full-set publisher evaluation with tools and context management. Text-only evaluations are shown separately.
Return to resultGLM 5.3 Flash · 55.3%Publisher-reported result · zai-org/GLM-5.3-Flash · Effort: Max · Humanity's Last Exam with tools, full set. Max effort; temperature 1, top_p 0.95, 163,840 maximum output tokens and 300,000-token context with context management. Judge: GPT-5.6 Luna Medium.
Publisher source - GPQA Diamond
Luna 90.47, GLM Flash 90.15 and DeepSeek Flash 0731 91.04 use exact Epoch model-version rows at Max. Qwen 91.7 is publisher-reported Flash-Next. MiMo has no matching verified result.
Return to resultNeurion Flash · 88.7%HydraLabs-reported result · Effort: Ultra
Qwen 3.8 Flash · 91.7%Publisher-reported result · Qwen/Qwen3.8-Flash-Next · Publisher result for Flash-Next, the basis of the Flash API. Evaluation harness, reasoning effort and exact API snapshot were not specified.
Publisher sourceGLM 5.3 Flash · 90.15%Independent evaluation by Epoch AI · glm-5.3-flash_max · Effort: Max · Evaluation date: 2026-08-26 · Exact model-version row from Epoch gpqa_diamond.csv; mean_score converted to a percentage and rounded to two decimals.
Epoch AI · Independent evaluationDeepSeek V4 Flash 0731 · 91.04%Independent evaluation by Epoch AI · deepseek-v4-flash-0731_max · Effort: Max · Evaluation date: 2026-08-02 · Exact model-version row from Epoch gpqa_diamond.csv; mean_score converted to a percentage and rounded to two decimals.
Epoch AI · Independent evaluationGPT-6 Luna · 90.47%Independent evaluation by Epoch AI · gpt-6-luna_max · Effort: Max · Evaluation date: 2026-09-22 · Exact model-version row from Epoch gpqa_diamond.csv; mean_score converted to a percentage and rounded to two decimals.
Epoch AI · Independent evaluation - Terminal-Bench Science 0.1
Vals upstream v0.1.0 configuration: 70 tasks, Mini-SWE-agent, single-run pass@1 and an eight-hour task allowance.
Return to resultMiMo V2.6 Flash · 4.29%Independent evaluation by Vals AI · xiaomi/mimo-v2.6-flash · Mini-SWE-agent · Terminal-Bench-Science 0.1, upstream v0.1.0; 70 tasks across five domains. Mini-SWE-agent; pass@1 from one run, with retries only for infrastructure or provider errors. The agent phase permits eight hours per task.
Vals AI · Independent evaluationGLM 5.3 Flash · 2.86%Independent evaluation by Vals AI · zai/glm-5.3-flash · Effort: Max · Mini-SWE-agent · Terminal-Bench-Science 0.1, upstream v0.1.0; 70 tasks across five domains. Mini-SWE-agent; pass@1 from one run, with retries only for infrastructure or provider errors. The agent phase permits eight hours per task.
Vals AI · Independent evaluationDeepSeek V4 Flash 0731 · 4.29%Independent evaluation by Vals AI · deepseek/deepseek-v4-flash-0731 · Effort: High · Mini-SWE-agent · Terminal-Bench-Science 0.1, upstream v0.1.0; 70 tasks across five domains. Mini-SWE-agent; pass@1 from one run, with retries only for infrastructure or provider errors. The agent phase permits eight hours per task.
Vals AI · Independent evaluation - CyberGym
Reported performance on CyberGym security tasks.
Return to resultNeurion Flash · 81.3%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
MiMo V2.6 Flash · 95.1%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Publisher table; run configuration not specified.
Publisher sourceDeepSeek V4 Flash 0731 · 76.7%Publisher-reported result · deepseek-ai/DeepSeek-V4-Flash-0731 · Publisher table; run configuration not specified.
Publisher source - SEC-Bench Pro
Reported performance on SEC-Bench Pro.
Return to resultNeurion Flash · 43.5%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
MiMo V2.6 Flash · 47.5%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · Publisher table; run configuration not specified.
Publisher sourceGPT-6 Luna · 34.2%Publisher-reported result · gpt-6-luna · OpenAI publisher evaluation of SEC-Bench Pro, May 2026 version covering V8 and SpiderMonkey. The benchmark contains 183 validated vulnerabilities. The card describes agent-assisted root-cause grading. Per-result reasoning effort, harness and run date were not specified.
Publisher source - ExploitGym
Reported performance on ExploitGym. No shared time budget or task scope is assumed.
Return to resultNeurion Flash · 12.3%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
MiMo V2.6 Flash · 6%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · The publisher table does not specify a time budget or userspace task scope.
Publisher sourceGPT-6 Luna · 11.6%Publisher-reported result · gpt-6-luna · OpenAI publisher evaluation of ExploitGym v1, reporting intended-exploit success. The benchmark contains 869 challenges across userspace, V8 and Linux kernel vulnerabilities. Runs remain offline with runtime package installation disabled; valid success requires remote execution and the intended vulnerability. Per-result effort, harness and token cap were not specified.
Publisher source - ExploitBench
A common metric definition is unavailable. Coverage and full-exploit success rates are distinct measures.
Return to resultNeurion Flash · 28.4%HydraLabs-reported result · Effort: Ultra · MelissaCode 0.1.0
MiMo V2.6 Flash · 25.3%Publisher-reported result · XiaomiMiMo/MiMo-V2.6-Flash-RL · The publisher reports ExploitBench without defining a coverage or full-exploit success metric.
Publisher source - FrontierMath Tier 4 (v2)
Exact Epoch model-version rows at Max. Raw mean_score values are converted to percentages and rounded to two decimals.
Return to resultGLM 5.3 Flash · 17.07%Independent evaluation by Epoch AI · glm-5.3-flash_max · Effort: Max · Evaluation date: 2026-08-27 · FrontierMath Tier 4 v2 at Max effort. The exact Epoch model-version row supplies the run date, raw mean and standard error. Percentage equals the raw mean multiplied by 100, rounded to two decimal places.
Epoch AI · Independent evaluationDeepSeek V4 Flash 0731 · 24.39%Independent evaluation by Epoch AI · deepseek-v4-flash-0731_max · Effort: Max · Evaluation date: 2026-08-02 · FrontierMath Tier 4 v2 at Max effort. The exact Epoch model-version row supplies the run date, raw mean and standard error. Percentage equals the raw mean multiplied by 100, rounded to two decimal places.
Epoch AI · Independent evaluationGPT-6 Luna · 56.1%Independent evaluation by Epoch AI · gpt-6-luna_max · Effort: Max · Evaluation date: 2026-09-22 · FrontierMath Tier 4 v2 at Max effort. The exact Epoch model-version row supplies the run date, raw mean and standard error. Percentage equals the raw mean multiplied by 100, rounded to two decimal places.
Epoch AI · Independent evaluation