AgentsAutonomy & tool use
Hermes Agent's built-in fallback, not one profile per model, is what keeps free Solar Mini 4 setups running through rate limits
A creator's Hermes Agent tutorial isolates each free model in its own profile, but Hermes already offers automatic provider fallback. Upstage's Solar Mini 4, free briefly on Nous Portal, scores well independently yet is verbose enough to erode its low token price.
AT Pulse research deskOctober 8, 2026 · 5 min read · AI-assisted analysisThe linked post gives each model its own Hermes profile. A profile keeps its own config, memory, sessions and skills, so this suits side-by-side testing, but profiles don't fail over. Hermes has a documented fallback_providers setting that switches to a backup provider and model mid-session after rate-limit or server errors and keeps the conversation history. It fires at most once per user message, and the backup model loses the cached-prompt discount on input tokens. [1] [3] [12] [4]
Solar Mini 4 is a proprietary mixture-of-experts model with 35B total and 3B active parameters. Nous Research says it's free on Nous Portal for two weeks but hasn't published an end date or a usage allowance. Free and paid versions appear next to each other in the model list, so users need to pick the :free variant to avoid being billed. [6] [5] [10] [3]
Artificial Analysis independently scores the model 24 on its Intelligence Index, against a median of 12 for comparably priced reasoning models. It also flags heavy verbosity: 370M output tokens on the index versus a 100M median. That makes cost per task a better budgeting guide than list price per token. The sources also disagree on speed: the Artificial Analysis page lists 76.5 tokens/s, below the median, while OrcaRouter's summary of the same evaluation cites 204. [8] [7]
Hermes documents built-in failover, and Artificial Analysis independently measured Solar Mini 4's score and heavy output.
Read the full assessment
Implication: teams relying on free models should configure fallback for unattended work and budget per task, since free access is temporary and its limits unpublished.
Executive brief
The linked post presents "one Hermes profile per model" as its way of coping when a free model hits a rate limit or loses free access. But Hermes Agent already has a built-in fallback that automatically switches to a backup provider and model mid-session after rate-limit or server errors (Hermes docs: Fallback Providers). The model behind the post is Upstage's Solar Mini 4, which Nous Research says is free on Nous Portal for two weeks (Nous Research on X). Its Artificial Analysis score of 24 is independently measured, but the same evaluation found it very verbose. That makes the price per million tokens a misleading guide to the real cost per task.
What changed and event timeline
Upstage ships Solar Mini 4
The model ID is
solar-mini4-260922: a 35B-parameter mixture-of-experts with 3B active, a 512K context window and up to 128K output tokens ().Artificial Analysis publishes its evaluation
It reports a score of 24 on Intelligence Index v4.3.2, against a median of 12 for comparably priced reasoning models (;).
Hermes adds Solar to its model pickers
A merged pull request adds
upstage/solar-mini4andsolar-pro4to the OpenRouter and Nous Portal pickers ().The free variant is confirmed callable
A community registry checked on this date lists
upstage/solar-mini4:freeamong Nous Portal's free models. It notes that Solar Pro 4 was removed from the free list on Oct 1 ().The linked post is published
The Reddit commentary describes setting up one profile per model in Hermes desktop, then upsells the author's paid community ().
Capabilities and access
- Model: Upstage Solar Mini 4 (
solar-mini4-260922). Strongest in Korean, also strong in English and Japanese. Training data runs to February 2026 (Upstage). - Weights: proprietary; weights are not public (Artificial Analysis).
- Paid price: $0.10 per million input tokens and $0.40 per million output tokens. OrcaRouter reports a 50% launch discount running to Oct 22 (OrcaRouter).
Read the full section
- Model: Upstage Solar Mini 4 (
solar-mini4-260922). Supports chat, reasoning, structured outputs and tool calling. Strongest in Korean, also strong in English and Japanese. Training data runs to February 2026 (Upstage). - Weights: proprietary; weights are not public (Artificial Analysis).
- Paid price: $0.10 per million input tokens and $0.40 per million output tokens. OrcaRouter reports a 50% launch discount running to Oct 22 (OrcaRouter).
- Free access: through the
:freevariant on Nous Portal. The free allowance is not published, and paid variants fall outside the free plan (registry).
Technical analysis for researchers and developers
- What a profile is: a separate Hermes home directory with its own
config.yaml,.env, memory, sessions, skills, cron jobs and state database (Profiles docs). - Built-in failover: the
fallback_providerssetting inconfig.yamlswitches provider and model in place, keeping conversation history. - Practical combination: profiles for testing models, fallback for keeping work running.
Read the full section
- What a profile is: a separate Hermes home directory with its own
config.yaml,.env, memory, sessions, skills, cron jobs and state database (Profiles docs). Profiles keep agents isolated, which makes side-by-side model comparisons fair. They do not provide failover. - Built-in failover: the
fallback_providerssetting inconfig.yamlswitches provider and model in place, keeping conversation history. It triggers on 429 or 5xx errors once retries run out, and immediately on 401, 403 or 404. Two limits: it fires at most once per user message, and the backup model has no cached prompt prefix, so the input-token discount is lost after a switch (Fallback docs). - Practical combination: profiles for testing models, fallback for keeping work running.
Claims and evidence
- Score of 24 on the Intelligence Index.
- The Artificial Analysis page lists 76.5 output tokens/s, below the 111.2 median. OrcaRouter's summary of the Artificial Analysis results cites 204 tokens/s instead (OrcaRouter).
- Upstage says 512K, the post says about 500K, and the Artificial Analysis page says 1.0M.
Read the full section
- Score of 24 on the Intelligence Index. Independent (Artificial Analysis). Nous repeats it and adds that it beats models with 10× the active parameters; that comparison is vendor-reported (Nous on X).
- "Fast." Contested. The video calls it fast (03:29). The Artificial Analysis page lists 76.5 output tokens/s, below the 111.2 median. OrcaRouter's summary of the Artificial Analysis results cites 204 tokens/s instead (OrcaRouter).
- Context window. Upstage says 512K, the post says about 500K, and the Artificial Analysis page says 1.0M.
- Free for two weeks. Vendor-reported (Nous on X; video 00:35). No end date is published.
- Found a news story the author had missed. Anecdotal; the item named in the auto-captions has no corroboration (03:39).
Context and prior work
Nous Portal's free list has changed from week to week: Laguna, Ling and LongCat 2.0 were added Sep 30 and Solar Pro 4 was dropped Oct 1 (registry). The post's list includes "Ling 3.1 Flash", which the registry does not show; it lists Ling 3.0 Flash variants instead.
Read the full section
Nous Portal's free list has changed from week to week: Laguna, Ling and LongCat 2.0 were added Sep 30 and Solar Pro 4 was dropped Oct 1 (registry). The post's list includes "Ling 3.1 Flash", which the registry does not show; it lists Ling 3.0 Flash variants instead. The video's auto-captions also contain errors, for example "Mistral" LongCat (01:55), where the model is Meituan's. Nous recently shipped a Profile Builder in the Hermes dashboard (CO/AI).
Limitations, safety and contested findings
- Evidence quality: the post is commentary from a seller of paid training, and it rests on one informal test.
- Not frontier-level: the presenter says so himself and warns about rate limits (03:20, 04:58).
- Cost per task: Artificial Analysis says the model is "very verbose": 370M output tokens on the index versus a 100M median, at about $0.36 per task (Artificial Analysis).
Read the full section
- Evidence quality: the post is commentary from a seller of paid training, and it rests on one informal test.
- Not frontier-level: the presenter says so himself and warns about rate limits (03:20, 04:58).
- Cost per task: Artificial Analysis says the model is "very verbose": 370M output tokens on the index versus a 100M median, at about $0.36 per task (Artificial Analysis).
- Privacy: Nous Portal data may be used for training unless Privacy Mode is enabled (registry).
Business and practitioner implications
- Make sure the
:freeentry is selected. - Compare models in separate profiles, and set
fallback_providersfor unattended work instead of switching by hand. - Budget per task, not per token. Verbose reasoning can wipe out low list prices once the free window closes.
Read the full section
- Check the variant. Make sure the
:freeentry is selected; the video shows a free and a paid Solar Mini 4 side by side in the same list (04:40). - Test, then fail over. Compare models in separate profiles, and set
fallback_providersfor unattended work instead of switching by hand. - Budget per task, not per token. Verbose reasoning can wipe out low list prices once the free window closes.
- Mind the data. Turn on Privacy Mode before sending client data through free endpoints.
The source trail.
Sources (13)
Why My Hermes Free Model Setup Uses One Profile Per Model
Feed excerpt only. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=DPbbSzUAYhM); automatic text may contain errors.
reddit.comSolar Mini 4: FREE Hermes Agent Model!
Related coverage; assess separately
reddit.comLinked video: Why My Hermes Free Model Setup Uses One Profile Per Model
linked video auto captions
www.youtube.com