Ministral-3-14B-Reasoning-MistralAI Hermes Benchmark Failure Report

Result: Failed / partial run

The model completed the initial prompt read and wrote a markdown planning report, but it did not produce the required benchmark HTML file. After a long generation period with no new file output, I interrupted the stalled finalization turn and asked Hermes to continue. The continuation then failed three times on the same LM Studio prompt-template error.

Model label
Ministral-3-14B-Reasoning-MistralAI
LM Studio identifier
mistralai/ministral-3-14b-reasoning
Hermes displayed model
ministral-3-14b-reasoning
Reasoning effort
high
Session key
20260527_055640_dfa588
Hermes session duration
15m 41s
Session messages
6 total: 2 user messages, 4 tool-call-related messages
Context at failure
31.5K / 78K tokens, about 40%

Outcome

Requirement Status Detail
Confirm active model Passed Hermes showed ministral-3-14b-reasoning; LM Studio showed mistralai/ministral-3-14b-reasoning.
Set reasoning to high Passed Hermes accepted /reasoning high and confirmed the setting.
Read benchmark prompt Passed Hermes read /Users/armanshawon/Documents/Benchmark/test-prompt.md.
Create required HTML artifact Failed No fresh hermes_local_model_agent_benchmark_submission.html or model-specific final HTML was created by Hermes for this run.
Recover after stall Failed A continuation prompt triggered repeated LM Studio template failures.

Timeline

Time Event
2026-05-27 05:56:40 +0600 Fresh Hermes session opened. Session key: 20260527_055640_dfa588.
2026-05-27 05:56:55 +0600 Benchmark instruction sent for mistralai/ministral-3-14b-reasoning.
Early run LM Studio moved through prompt processing and active generation. Hermes context reached about 21.3K / 78K after prompt read.
About 3 minutes elapsed Hermes read the benchmark prompt file successfully.
2026-05-27 06:05 +0600 Hermes wrote Ministral-3-14B-Reasoning-MistralAI-benchmark-report.md, but it was only a planning report, not the required benchmark output.
About 14m 46s elapsed Generation had continued for 368.7 seconds after the last file write without producing the final HTML. I interrupted the API call.
2026-05-27 06:12:05 +0600 Continuation attempt failed after three built-in Hermes retries with the same Jinja prompt-template error.
2026-05-27 06:12:31 +0600 Failure report prepared after closing the Hermes session.

Tool Call Audit

Order Hermes tool Target Observed result
1 read_file /Users/armanshawon/Documents/Benchmark/test-prompt.md Completed. TUI showed about 1.1s.
2 write_file /Users/armanshawon/Documents/Benchmark/Ministral-3-14B-Reasoning-MistralAI-benchmark-report.md Completed. 1,943 bytes written. TUI showed about 0.8s.
3 Continuation model call Attempt to produce Ministral-3-14B-Reasoning-MistralAI-high.html Failed before any tool call because LM Studio rejected the prompt template.

Error Details

The original long generation did not show a provider error; LM Studio reported GENERATING while no new file appeared. After interruption and continuation, Hermes recorded three API failures from the local provider.

Provider: custom
Model: mistralai/ministral-3-14b-reasoning
Endpoint: http://127.0.0.1:1234/v1
Error: Error rendering prompt with jinja template: "After the optional system message, conversation roles must alternate user and assistant roles except for tool calls and results."
Retries: 3/3 exhausted
Request debug dump: /Users/armanshawon/.hermes/sessions/request_dump_20260527_055640_dfa588_20260527_061205_530413.json

Partial Artifact

Hermes created this partial markdown report: /Users/armanshawon/Documents/Benchmark/Ministral-3-14B-Reasoning-MistralAI-benchmark-report.md. The markdown explicitly says the HTML file still needed to be generated, so it is not a valid benchmark completion artifact.

Judgment

This model run should be recorded as a failure for the benchmark. It demonstrated basic tool use by reading the prompt and writing a small markdown file, but it failed the required deliverable and could not recover after interruption because of a model prompt-template incompatibility in LM Studio.