ModelsArchitectures & capability
Upstage's Solar Mini 4 Free for Two Weeks on Nous Portal for Hermes Agent Users
Upstage's Solar Mini 4 is a 35B mixture-of-experts model that uses 3B parameters per token. It is free for about two weeks on Nous Portal for Hermes Agent users. Benchmarking finds it slow and wordy, though it scores well on reasoning.
QuietHow it’s measured
YouTube video · 151 views · 1 like
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
AT Pulse research deskPublished · 5 min read · AI-assisted analysisNous announced two weeks of free Solar Mini 4 access on October 5, which puts the end at roughly October 19. The free tier only applies if users pick the variant marked free, upstage/solar-mini4:free. A paid listing that looks almost the same draws on credits instead. Nous used the same naming rule in an earlier free promotion for Nemotron 3 Ultra. [8] [10] [3] [5]
Artificial Analysis, the only independent benchmark cited, measured 64.4 output tokens per second and 1.79 seconds to the first token, and rated that slow. That contradicts the promotional claim that the model is fast. Its Intelligence Index score was 24, against a peer median of 13. It also produced 370M output tokens during the evaluation, against a median of 100M, which can cancel out its low per-token price in agent loops. [4] [6] [3]
The video's most useful advice is about operations. Keep a separate Hermes profile for each model so you can switch in one click when a rate limit hits or a free window ends. Send bulk work like research and sorting to light models, and send harder reasoning to stronger ones. [3]
By default, Nous Portal may store inference data and use it for training unless Privacy Mode is turned on. Nous also does not publish usage allowances for the free tier. Neither the linked post nor the video mentions either point. [10]
The evidence shows a model that is cheap per token but slow and wordy. For teams, that means free promotions suit low-stakes bulk agent work.
Read the full assessment
Keep fallback models ready, turn on Privacy Mode, and measure response time and output length on your own workloads.
Executive brief
The video sells Solar Mini 4 as "fast", but the only independent benchmark rates it slow. Artificial Analysis measured 64.4 output tokens per second and called that notably slow. The same benchmark found it unusually wordy, even though its reasoning score beat most comparable models. The linked post describes a free two-week offer of Upstage's Solar Mini 4 on Nous Portal, used through Nous Research's Hermes Agent. Its real advice is sensible: keep backup model profiles so a rate limit or an ended promotion doesn't stop your work. The post and video are promotional, though, and they leave out how long the free window lasts and what happens to your data.
What changed and event timeline
Nous runs an earlier free-model promotion
Nemotron 3 Ultra was free on Nous Portal for two weeks, but only if users picked the
:freevariant. The Solar Mini 4 offer follows the same pattern.Upstage releases Solar Mini 4
It launched as model
solar-mini4-260922, a 35B mixture-of-experts model with 3B parameters active at a time.More detail
Solar Mini 4 becomes free on Nous Portal
Nous announced two weeks of free access and pointed to the model's Artificial Analysis score of 24.
A community list confirms the free model ID
It lists
upstage/solar-mini4:freeamong eight free Nous Portal models. Nous does not publish how much free usage each account gets.The Reddit post and video appear
The video shows three setup routes, warns about the paid look-alike listing, and promotes a paid community.
More detail
Capabilities and access
- Model: Solar Mini 4, ID
solar-mini4-260922(aliassolar-mini4). Text in and out, in English, Korean and Japanese. Knowledge runs to February 2026. Upstage docs - Context: 512K tokens in and up to 128K out, per Upstage. Artificial Analysis lists 1M instead, so the figures conflict.
- Price: List price is $0.10 per million input tokens and $0.40 per million output. OrcaRouter reports half price until October 23.
Read the full section
- Model: Solar Mini 4, ID
solar-mini4-260922(aliassolar-mini4). Text in and out, in English, Korean and Japanese. Supports tool calling, structured outputs and a reasoning mode. Knowledge runs to February 2026. Upstage docs - Context: 512K tokens in and up to 128K out, per Upstage. Artificial Analysis lists 1M instead, so the figures conflict.
- Price: List price is $0.10 per million input tokens and $0.40 per million output. OrcaRouter reports half price until October 23.
- Free access: Through Hermes Cloud, Hermes Desktop or the
hermes dashboardcommand, choosing the variant marked FREE. AI Success Lab - Weights: Not public; the model is proprietary.
Technical analysis for researchers and developers
- Architecture: A sparse mixture-of-experts model that uses roughly 9% of its parameters for each token. That design aims for low cost per token.
- Verbosity: The model produced 370M output tokens during the benchmark run, against a median of 100M.
- Integration: Nous Portal offers an OpenAI-compatible API. Free access depends on the exact
:freesuffix. awesome-free-ai-coding
Read the full section
- Architecture: A sparse mixture-of-experts model that uses roughly 9% of its parameters for each token. That design aims for low cost per token. It does not guarantee fast responses: Artificial Analysis measured 64.4 tokens/s, with 1.79 seconds to the first token. Artificial Analysis
- Verbosity: The model produced 370M output tokens during the benchmark run, against a median of 100M. For agent loops, that inflates cost and response time per task.
- Integration: Nous Portal offers an OpenAI-compatible API. Free access depends on the exact
:freesuffix. awesome-free-ai-coding - Reproducibility: The video's only "test" is one prompt summarising news, with no comparison or scoring (04:00).
Claims and evidence
- 35B total / 3B active parameters: Matches Upstage's own documentation (video 01:03; Upstage).
- "It's fast" (00:06): Contradicted by the throughput Artificial Analysis measured.
- Intelligence Index of 24, "above models with 10x the active parameters": Nous makes this comparison (X post). The "10x" framing is Nous's own.
Read the full section
- 35B total / 3B active parameters: Matches Upstage's own documentation (video 01:03; Upstage).
- "It's fast" (00:06): Contradicted by the throughput Artificial Analysis measured.
- Intelligence Index of 24, "above models with 10x the active parameters": Nous makes this comparison (X post). It rests on Artificial Analysis's score, where the peer median is 13. The "10x" framing is Nous's own.
- "Holds up way better than you'd guess" (01:35): Opinion based on one anecdote. One day after launch, OrcaRouter said no independent evaluation existed yet.
- Paid and free listings look alike (05:14): Consistent with the
:freenaming rule Nous documented for its earlier promotion.
Context and prior work
- Nous Portal has run time-limited free-model promotions before: Nemotron 3 Ultra in June, with NVIDIA and Nebius. Hermes docs
- Other free models on the portal are Step 3.7 Flash, Laguna S/XS 2.1, Ling 3.0 Flash variants and LongCat.
- Solar Pro 4, the larger model in the same family, shares the 512K/128K context spec. OrcaRouter
Read the full section
- Nous Portal has run time-limited free-model promotions before: Nemotron 3 Ultra in June, with NVIDIA and Nebius. Hermes docs
- Other free models on the portal are Step 3.7 Flash, Laguna S/XS 2.1, Ling 3.0 Flash variants and LongCat. These are the "swap-in" options the video refers to. awesome-free-ai-coding
- Solar Pro 4, the larger model in the same family, shares the 512K/128K context spec. OrcaRouter
Limitations, safety and contested findings
- Data use: By default, Nous Portal may store inference data and use it for training. Neither the video nor the post mentions this.
- Free window: About two weeks from October 5, so roughly October 19 by calculation.
- Undisclosed limits: Nous does not publish free-tier usage allowances.
Read the full section
- Data use: By default, Nous Portal may store inference data and use it for training. Users have to switch on Privacy Mode to stop that. awesome-free-ai-coding Neither the video nor the post mentions this.
- Free window: About two weeks from October 5, so roughly October 19 by calculation. The video only says the model is free "right now" (02:12).
- Undisclosed limits: Nous does not publish free-tier usage allowances.
- Conflicting specs: Context length is listed as 512K by Upstage and 1M by Artificial Analysis.
- Who is promoting it: The video presenter is an AI avatar. The setup guides for this story mostly come from websites belonging to the same paid community.
Business and practitioner implications
- Design for model churn: Keep one profile per model, as the video suggests (07:00). Free promotions are temporary marketing.
- Route by task: Send cheap bulk work (extraction, sorting, simple tool calls) to small models like this one, and harder reasoning to stronger models (07:15).
- Before sending business data through a free tier, enable Privacy Mode, pin the exact
:freemodel ID, and budget for the paid price once the promotion ends.
Read the full section
- Design for model churn: Keep one profile per model, as the video suggests (07:00). Free promotions are temporary marketing.
- Route by task: Send cheap bulk work (extraction, sorting, simple tool calls) to small models like this one, and harder reasoning to stronger models (07:15).
- Before sending business data through a free tier, enable Privacy Mode, pin the exact
:freemodel ID, and budget for the paid price once the promotion ends. - Test it yourself: Measure response time and output length on your own workloads before relying on it. The model's wordiness can cancel out its low per-token price.
The source trail.
Sources (10)
How I Run Hermes Agent for Free and Swap Models in One Click
Feed excerpt only. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=7trYKCvueao); automatic text may contain errors.
reddit.comRun Hermes Agent for FREE With Solar Mini 4
Related coverage; assess separately
reddit.comLinked video: How I Run Hermes Agent for Free and Swap Models in One Click
linked video auto captions
www.youtube.com