DeepSeek V4 Flash vs Amazon Nova Lite

WTF, I could just use the expensive ai models of opencode go for planning, writing specs and then use opencode zen deepseek v4 flash max for implementation. I am loving this opencode, loving the freebies

Just wanted to share that I used u/LegacyRemaster slightly modified (Q4\_K\_M conversion support) DeepSeek V4 [CUDA repo](https://github.com/Fringe210/llama.cpp-deepseek-v4-flash-cuda) (based on u/antirez [work](https://github.com/antirez/llama.cpp-deepseek-v4-flash)) to convert and run Q4\_K\_M [DeepSeek V4 Pro](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro) on my Epyc workstation (Genoa 9374F, 12 x 96GB RAM, single RTX PRO 6000 Max-Q) and it worked right from the start:

(base) phm@epyc:~/projects/llama.cpp-deepseek-v4-flash-cuda/build-cuda$ ./bin/llama-cli -m ../models/DeepSeek-V4-Pro-Q4_K_M.gguf --no-repack -ub 128 --chat-template-file ../models/templates/deepseek-ai-DeepSeek-V3.2.jinja
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 97247 MiB):
Device 0: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, compute capability 12.0, VMM: yes, VRAM: 97247 MiB

Loading model...

▄▄ ▄▄
██ ██
██ ██ ▀▀█▄ ███▄███▄ ▀▀█▄ ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██ ██ ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
██ ██
▀▀ ▀▀

build : b8936-44c7b01de
model : DeepSeek-V4-Pro-Q4_K_M.gguf
modalities : text

available commands:
/exit or Ctrl+C stop or exit
/regen regenerate the last response
/clear clear the chat history
/read <file> add a text file
/glob <pattern> add text files using globbing pattern

> who are you?

[Start thinking]
Okay, the user is asking "who are you?" This is a simple, introductory question. I need to introduce myself clearly and warmly. I should state my name, creator, and key features that are most relevant to a new user. I can mention that I'm free, my context window, knowledge cutoff, file support, and availability on web and app. I'll end with an open invitation for further questions to keep the conversation going.
[End thinking]

Hi there! I'm DeepSeek, an AI assistant created by the Chinese company DeepSeek (深度求索). I'm here to help you with questions, creative tasks, problem-solving, and pretty much anything you're curious about!

Here's a bit about me:
- **Free to use** - no charges for chatting with me
- **1M context window** - I can handle huge amounts of text at once (like entire book trilogies!)
- **Knowledge cutoff: May 2025** - I'm reasonably up-to-date
- **File upload support** - I can read text from images, PDFs, Word docs, Excel files, and more
- **Web search capability** - though you need to manually enable it via the search button
- **Available on web and mobile app** - with voice input support on the app

I'm a pure text-based model, so I can't "see" images directly, but I can read any text in uploaded files. I aim to be warm, helpful, and detailed in my responses.

What can I help you with today? 😊

[ Prompt: 12.2 t/s | Generation: 8.6 t/s ]

> /exit

Exiting...
common_memory_breakdown_print: | memory breakdown [MiB] | total free self model context compute unaccounted |
common_memory_breakdown_print: | - CUDA0 (RTX PRO 6000 Blackwell Max-Q Workstation Edition) | 97247 = 4022 + ( 92472 = 87766 + 84 + 4621) + 753 |
common_memory_breakdown_print: | - Host | 793994 = 793954 + 0 + 39 |
~llama_context: CUDA_Host compute buffer size of 39.1719 MiB, does not match expectation of 15.3535 MiB

The model file is 859GB.

Update: ran some lineage-bench prompts to see if the model has healthy brain and no problems so far.

Hi all.

I've been testing out different models and providers to see what is the best bang for buck you can get for around $20 if you are not running local models.

I have a Hermes agent running on a VM with 6GB RAM, which I got for an absolute steal of $45 per year (check out the LowEndTalk forum for cheap VPS deals). I use it mainly to maintain a dashboard that does the following:

* Gather news on specific topics from various sources. It then curates them to see if they align with my interests (eg. no sensasionalist crap), summarizes and deduplicates articles.
* Check the latest benchmarks on different models
* Scrape my favourite webcomics from Instagram, RSS feeds, Bluesky, whatever, so they are all in one place.

It also maintains the VPS, so I have it install docker containers for stuff I want, like Mealie or whatever.

Lastly, I synced my Obsidian vault where I keep a list of people with birthdays, notes etc. So it can remind me who's birthday it is and what I can buy for them, or other stuff like that. My Obsidian is also where it keeps track of my health stuff. Diet, gym log, etc.

So, I've been playing around with the following providers. In all cases except Codex and OpenRouter, I used Kimi K2.6 as my main model, and usually tried Gemma4 for some of the tools and auxiliary models:

* Ollama Cloud - $20 per month
* OpenCode Go - $10 per month
* NanoGPT - $12 per month (I think you can get $8 if you find a ref link)
* OpenAI Codex - $20
* OpenRouter - Free Models only

Here are my findings.

# Ollama Cloud

Very stable. Charges per GPU hours instead of tokens, so as models get more efficient, you actually gain mode usage. Some people say it's a bit slow, but in my experience it was never slow enough to be problematic.

I actually had a hard time hitting my usage limits. I had to run my Hermes Agent, as well as 2 pretty big coding tasks simultaneously before I hit my 5 hour window limit, and this only happened once. The rest of the time, I barely cracked 25%. For Hermes alone, you will likely never hit that limit.

Cons, are that you are limited to 3 concurrent connections. Meaning, my example of 2 coding cases and Hermes was pushing it. If I had to chat to Hermes and a cron job fired that used a model, it errored out because I went over the limit of 3 connections. This is something to keep in mind for people running multiple agents or lots of cron jobs and such.

# OpenCode Go

I felt like this was ever so slightly less stable than Ollama, but not enough to be a problem or to stay away from it. Speed was fine, I honestly didn't feel much of a difference between OpenCode and Ollama. You pay $10 per month, and essentially get $60 worth of credits.

One might think $60 credits is not much, but whether it is an efficiency thing or just the fact that we aren't paying Anthropic pricing, it stretched very far. I never hit my limits. Just like Ollama, on average usage I barely got to 25-30% weekly. Unlike Ollama, you don't have concurrency limits.

The con for me is that it didn't have the model I wanted for tool calls, Gemma 4. They don't have that on here. They have DeepSeek which is cheap and fast, but Gemma 4 is cheap, fast AND multimodal. Useful for curating news articles or webcomics.

# NanoGPT

This one seemed sketchy AF at first. It's clearly meant for a specific crowd. It has a ton uncensored text models included in the sub, as well as uncensored image models (Qwen Image and Z Image Turbo) with 100 free image generations per day. They allow you to load up with crypto (or visa if you don't have crypto) and sign in with only a passkey, no need to enter an email or anything, allowing for a degree of anonymity.

Kimi on this one was VERY verbose. It thought a lot, and then would output that as messages in Telegram, meaning the chat context grew very, very fast and had to compress every couple of messages. They had Gemma 4 though (a bunch of variations), and using them for tool calls worked fine. Of this list, NanoGPT had the most models available on the sub. Usage limits seemed a lot lower than Ollama and OpenCode. Also worth noting, since the model naming on this one is a bit weird, if you are relying on your main model to maintain it's own config, you need to give it the *exact* model you want to use. If you just tell it to use "Gemma 4" then high chance it will take the one not in your sub and complain about you needing to top up credits first.

# Codex

Currently testing. Ran it for a day and weekly usage is already at 30%. Didn't even push it that hard. Using GPT 5.5 on it. It feels like it is running an excessive number of tool calls whenever I give it a task. Doing random searches, terminal commands, notes, etc. I'll see if I hit my weekly in 3 days or not. I probably will.

# OpenRouter

The standard free models are extremely unreliable and often hit rate limits. However they also frequently have preview models that work very nicely for a week or 3, and are worth at the very least using for tool calls. They recently had Tencent Hy3 for free which even now is topping the LLM Leaderboard on OpenRouter. It is very much worth having an OR API key in your back pocket that you can plug into an auxiliary function or some cron jobs to save usage when things like this happen.

# Honorable Mention

**Nous Portal** \- You pay $20, you get $22 credits. Not a lot of savings. However they do have some free models from time to time as well. Right now they have Step 3.5 Flash and Deepseek V4 Flash for free. Need to top up your wallet before you can use them though. Like OpenRouter, worth having a key in your back pocket for the occasional freebie.

# My plan going forward

Once this month's codex runs out, I think I will likely stick with **OpenCode Go + NanoGPT**. I will use OpenCode Go for my main model, profiles, and maybe a bit of coding, and NanoGPT for auxiliary models and free image generation. I am paying $8 per month for Nano instead of $12, not sure how I got that discount, think it was an affiliate link probably. This means, my total setup will be **$18 per month** (or $22 if you don't get a discount) and I have access to a TON of models. I then still have some credits in Nous Portal and OpenRouter on the off chance I need something very niche.

The hype posts make it sound perfect. "INSANE (FREE!)." "makes $2,000/mo cloud stacks obsolete." "The gap is just 30 minutes of setup."

I've been running this stack for a few weeks now. It IS genuinely good. But nobody talks about the tradeoffs honestly. So here's the full picture.

**The setup (it's real and it's actually $0):**

Cloud route (fastest to try, still $0):

bash

ollama launch openclaw --model deepseek-v4-flash:cloud

One command. installs ollama, pulls the model, configures openclaw, launches the gateway. connect telegram. your agent is live.

Fully local route (after you have the hardware):

bash

ollama pull deepseek-r1:14b
# then configure openclaw to use the local ollama provider

Point openclaw at it with `api: "ollama"` And everything runs on your machine. data never leaves your network. no API keys. no subscriptions. genuinely $0 forever.

For the V4 Flash cloud route through Ollama, the model runs on Ollama's US-hosted servers. still free. still no API key needed. but your prompts do leave your machine, they just go to Ollama's infrastructure instead of DeepSeek's directly.

**What you gain (the real stuff):**

Privacy. Your data stays on your machine (fully local route) or at minimum stays within US infrastructure (ollama cloud route). nothing goes to Anthropic, OpenAI, or DeepSeek's Beijing servers.

Zero ongoing cost. No per-token billing. no subscription. No surprise $350 bills from a runaway cron job. The worst case is your electricity bill goes up.

No provider dependency. Anthropic can't ban your subscription. OpenAI can't change their pricing. DeepSeek can't rate limit you. Your agent runs whether the internet is on or off (fully local only).

DeepSeek V4 is genuinely capable. 1M token context window. mixture-of-experts architecture (1.6 trillion total parameters, 49 billion active per token for V4 Pro). strong at coding, reasoning, and agentic tasks. This isn't a toy model.

**Now here's what you actually lose:**

**Speed.** This is the big one nobody mentions in the hype posts. Local inference on consumer hardware is noticeably slower than cloud APIs. On a 16GB GPU running deepseek-r1:14b, expect maybe 15-25 tokens per second. Claude Sonnet on API gives you 120 tokens per second. You feel the difference on every single interaction. CPU-only setups are borderline unusable for agent work.

**Raw capability ceiling.** DeepSeek V4 Flash is excellent. but it's not Opus 4.7 or GPT-5.5 on the absolute hardest tasks. complex multi-step reasoning, nuanced creative work, flawless error recovery during tool chains. The gap is real on the top 10% of difficulty. for the other 90%? genuinely comparable.

**Hardware barrier.** The hype posts forget to mention you need actual hardware.

8GB VRAM: qwen 7b or deepseek-r1:1.5b. functional but limited.

16GB VRAM: deepseek-r1:14b. good enough for most agent tasks. the sweet spot for most people.

24GB+ VRAM: deepseek-r1:32b or V4 Flash quantized. best local experience. Requires a serious GPU or mac with unified memory.

V4 Pro locally? forget it unless you have a mac studio with 128GB+ unified memory. not happening on consumer hardware.

if you don't have 16GB+ VRAM, the fully local path is frustrating. use the ollama cloud route instead (still free, just not fully local).

**Reliability.** cloud APIs have teams monitoring uptime, handling failures, scaling capacity. your local setup has you. if ollama crashes at 3am, your morning briefing doesn't arrive. if your GPU overheats, your agent dies. if a power outage hits, everything stops. you are the sysadmin, the devops team, and the on-call engineer. all at once.

**Tool calling consistency.** local models are flakier on tool calls than cloud models. they'll occasionally skip a step in a multi-tool chain, hallucinate a tool result, or say "done!" when nothing happened. the smaller the model, the worse this gets. deepseek-r1:14b handles simple tool chains fine. complex 5+ step workflows get shaky.

**Setup and maintenance.** "30 minutes of setup" is optimistic. if everything works first try, maybe. but model downloads take time (14b is \~9GB, 32b is \~20GB). quantization issues happen. ollama config quirks appear. context limits in practice don't always match specs. updates aren't automatic. you're on bleeding edge with occasional bugs.

**The honest assessment:**

This stack is legitimately transformative for three types of people:

privacy-focused users who won't send data to cloud providers under any circumstances. The fully local route is the real deal. nothing leaves your machine.

tinkerers who enjoy the process of optimizing and maintaining their own setup. if debugging ollama configs at midnight sounds fun to you, this is your stack.

budget-constrained users who have the hardware but not the monthly budget. if you have a decent GPU sitting idle, this is free compute you're already paying for.

for everyone else? honestly, a hybrid setup makes more sense. run deepseek locally for routine daily tasks (briefings, simple research, drafts). fall back to a cloud API for the 10% of tasks that need frontier reasoning. your local setup handles the volume. the cloud handles the hard stuff.

**The one thing the hype posts get right:**

The gap between local and cloud is closing fast. A year ago, running an AI agent locally was a joke. Today, DeepSeek V4 Flash through ollama genuinely rivals cloud offerings for most daily agent use cases. a year from now, the gap might not matter for anyone.

But today, in May 2026, "fully local $0 agent" comes with real tradeoffs. knowing them upfront is the difference between a setup that lasts and one you abandon after a frustrating weekend.

If you're going to try it, start with the Ollama cloud route:

bash

# zero download, zero config, free
ollama launch openclaw --model deepseek-v4-flash:cloud

See if agent workflows are useful to you at all before investing in local hardware and fully-offline setup.

And if you don't want to manage any of this, there are managed platforms with free tiers that handle the infrastructure.

My customers have confidential data, they won't even use AWS.

I've been trying to solve this problem for them and they are more than fine with buying an on-premise device for Local LLMs + AI Agents.

Up until today, I have been extremely dissapointed with every model not named Opus.

However, Deepseek 4 Flash is doing near-Opus level performance. This is something I can actually use.

Upon this whole process things I dont understand:

>How are Qwen 35b people are using it? Not even sonnet can do the job.

>Do Mac users just say they are using local LLMs but not actually? That stuff is unbelievably slow. Heck, even with NVIDIA GPUs, it can be a bit frustrating when doing 1M tokens.

Anyway, thanks China for the free LLM. Not sure what they get out of it, I'm running it locally.

*Covers 2.5 weeks of discussion (Apr 30 – May 17, 2026). 32 threads analyzed. Split between Local and Cloud models, grouped by use case. Knowledge tables at the end.*

---

# LOCAL MODELS

Models that run on your own hardware via Ollama, LM Studio, or similar. Free to run — cost is your GPU/RAM.

---

## Qwen 3.6 (27B / 35B)

**Use: Community favorite — local self-hosted primary**
The most popular local model. Runs on everything from 8GB GPUs to 128GB RAM machines. The 27B variant is the sweet spot; 35B-a3b is the budget option.

- u/mrgreatheart: "I've been running Qwen3.6-27B-Q6_K for a while and it's fantastic." Uses the AEON uncensored variant. ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))
- u/fuchelio: "I use local Qwen 3.6 27B in full precision as the backend for a knowledge base system" ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/Thickdickmick87: "I'm finding qwen3.6-35b-a3b is pretty adequate. Running it locally on 8gb 3070" ([reddit](https://reddit.com/r/hermesagent/comments/1tfak3b/creating_obsidian_knowledge_base/))
- u/Express_Nebula_6128: "I've been using mostly qwen3.6 35b a3b running on my m4 max" ([reddit](https://reddit.com/r/hermesagent/comments/1teqzjg/questionconcern_are_you_using_nous_or_other/))
- u/Britbong1492 uses a routing system: "about 95% is done on a local qwen3.6:35b-A3b on my M4 Pro" with cloud fallback for hard tasks ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))

**Use: Benchmarking / optimization**
- Thread by u/Benchmarking the b9200 update: optimizing Qwen 3.6 27B multi-token prediction for Hermes Agent ([reddit](https://reddit.com/r/hermesagent/comments/1tg6vft/benchmarking_the_new_b9200_update/))

**Use: Knowledge base + Obsidian**
- u/JBManos: "I run qwen3.5-122b-a10b and it's having trouble with some obsidian tasks" — suggesting qwen3.6-27b as alternative ([reddit](https://reddit.com/r/hermesagent/comments/1tfak3b/creating_obsidian_knowledge_base/))

---

## Qwen 3.5

**Use: Larger local models for complex tasks**
- u/JBManos runs qwen3.5-122b-a10b locally — reports it struggles with some Obsidian tasks, suggesting smaller Qwen 3.6 may be better for certain workflows ([reddit](https://reddit.com/r/hermesagent/comments/1tfak3b/creating_obsidian_knowledge_base/))
- u/TexBluBoy uses Qwen 3.5 on a GMKtec EVO-X2 with AMD Ryzen AI Max+ 395 ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))
- u/krishna2910-amd asks about "qwen 3.5+ models" on local hardware ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))

**Use: Entry-level local (not recommended)**
- u/SecretSpace2 asks: "Is it not worth using lower tier models like Qwen3.5-9B?" — community advises against it for agent tasks ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))

---

## Gemma 4

**Use: Creative / SillyTavern (not recommended for Hermes)**
Mixed reviews. Some users like it for creative tasks, but consensus is it struggles with Hermes agent workflows.

- u/BehindUAll: "Don't use Gemma4 locally or using cloud API because it's horrible in Hermes." ([reddit](https://reddit.com/r/hermesagent/comments/1tfak3b/creating_obsidian_knowledge_base/))
- u/kunjukundi pushes back: "the model is downstream of the bigger issue: you're making Gemma do PDF parsing" — context matters ([reddit](https://reddit.com/r/hermesagent/comments/1tfak3b/creating_obsidian_knowledge_base/))
- u/PSyCHoHaMSTeRza: "the most common use case for it is SillyTavern" — niche creative use ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- u/ButterflyEconomist moved from cloud to local Gemma after learning about Hermes ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))
- u/Rootshot getting 128GB DDR5 for Geekom A9 Max to run Gemma properly ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))

---

## Llama 4 Maverick

**Use: Large local model for capable hardware**
- u/ButterflyEconomist mentions Llama 4 Maverick for running large models locally — requires beefy hardware (128GB+ RAM setups) ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))
- u/Rootshot is "getting 128gb DDR5 delivered later today" for Geekom A9 Max specifically to run large local models like Llama 4 ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))

---

## GLM (Zhipu)

**Use: Sweet-spot local model**
- u/itssethc: "GLM 5.1 is a nice sweet spot" — balanced between size and capability ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/TralfamadorianNode uses GLM alongside Qwen on Dappnode Next subscription ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- u/Present_Kitchen_9739 compares GLM to Haiku for agent tasks ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))

---

## Phi-4 Mini

**Use: Auxiliary / helper model**
- Not discussed as a primary model in threads, but referenced in Hermes profile configurations as the default auxiliary model for local setups. Handles sub-tasks like classification and summarization alongside the main model. Commonly paired with Qwen and Gemma local profiles.

---

# CLOUD / API MODELS

Models accessed via API or subscription. Pay per token or flat monthly rate.

---

## DeepSeek R1 / V4

**Use: Budget daily driver**
The go-to for cost-conscious users running heavy workloads. Community consensus: use the native DeepSeek provider for the best discounts — OpenRouter routing adds overhead and misses caching.

- u/mixxoh recommends DeepSeek V4 directly: "At most $1 a day" ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- u/SelectionCalm70: "just use deepseek v4 flash" as the default advice for newcomers ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/renoturx has "used nothing but free models from nous portal and openrouter" including DeepSeek — "been pretty ok" ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))
- u/cpatr922 notes Nous has free DeepSeek Flash access: "it is crazy experience" ([reddit](https://reddit.com/r/hermesagent/comments/1tf3f2f/my_estimated_tokens_cost_saving_in_a_month/))

**Use: Provider routing**
- u/torrso (PSA): "If you use deepseek, use the deepseek provider only. The others don't discount cached tokens." ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))
- u/verkavo asks which provider is best for DeepSeek — thread has routing recommendations ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- u/EconomyPhotograph927 had to tweak security settings so OpenRouter would let DeepSeek through ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))

**Use: Cost optimization**
- u/JordanPetterPans: "DeepSeek v4 has been just okayy" — adequate but not exceptional for their use case ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))

---

## Minimax

**Use: Token-plan budget model**
Minimax's token plan is popular among heavy users who want flat-rate pricing. Sentiment is mixed — cheap and uncensored but inconsistent quality.

- u/vandalieu_zakkart: "i am just happy with my minimax token plan. it's not the smartest but being virtually unlimited is nice" ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- u/yayita2500: "I use minimax token plan as main model but all my scripts are ready for minimax 2.5" ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/itsdodobitch: "Minimax with the token plan for me, but its quite dumb lately" ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/LouVillain: "Minimax-M2.1 via huggingface right now. I've been on almost all the frontier providers" ([reddit](https://reddit.com/r/hermesagent/comments/1tek5sw/advice_on_model/))
- u/kawasaki500 uses Minimax for "heavy AI code and usage, not worry about token" ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- Thread: Hermes & Minimax 2.5 problems — troubleshooting compatibility issues ([reddit](https://reddit.com/r/hermesagent/comments/1tg5968/hermes_minimax_25_problems/))

**Use: GPT-5.4 mini replacement**
- u/Immediate_Let_4946: "gpt 5.4 mini can be replaced with mini max. Just the darn gpt5.5 and Claude Sonnet are superior" ([reddit](https://reddit.com/r/hermesagent/comments/1tf3f2f/my_estimated_tokens_cost_saving_in_a_month/))

---

## MiMo (Xiaomi)

**Use: Token-plan alternative to Minimax**
MiMo is emerging as a Minimax competitor with aggressive pricing on token plans.

- u/Ok_Firefighter3363: "spend 6usd more (16usd) take mimo v2 pro 200 mil token plan. you will know the difference" ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))
- u/kawasaki500 uses Minimax + MiMo combo for heavy code usage ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- u/francxsim asks: "How is the Mimo 2.5 Pro experience switching from Minimax?" ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))

---

## Kimi K2.6 (Moonshot)

**Use: Coding and software development**
Kimi K2.6 is gaining traction as a Claude Sonnet alternative for coding at lower cost.

- u/8bit64k: "I've been using Kimi K2.6 nearly 100%. My use case right now is software development." ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/mf-mj asks about satisfaction and latency with Kimi K2.6 ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/bigdawg0420: "you're better off doing kimi k2.6 from the opencode go subscription" vs OpenAI sub ([reddit](https://reddit.com/r/hermesagent/comments/1tg0uyy/what_about_using_hermes_with_my_openai/))
- u/wtfzambo tried Kimi alongside Gemma4: "What's special about Gemma4? I tried for a bit but w.r.t. Kimi..." — implying Kimi is better ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- u/Thomas-Lore: Kimi "is close to Sonnet but only the largest models" you won't be able to run locally ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))

**Use: Writing**
- u/RawFreakCalm: "Kimi is a good writer like Claude. Gpt is an awful writer without really good system prompts." ([reddit](https://reddit.com/r/hermesagent/comments/1tfm0qd/model_selection_cold_outbound_email_with/))

---

## GPT-5.4 (OpenAI)

**Use: Subscription-based coding**
- u/dalemugford: "codex through sub as your main model, and o3 for hard reasoning" — using OpenAI subscription for cost-effective access ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- Thread asking about using Hermes with OpenAI subscription and gpt 5.4 mini ([reddit](https://reddit.com/r/hermesagent/comments/1tg0uyy/what_about_using_hermes_with_my_openai/))
- u/HobokenChickens rolled their own setup with GPT-5.4 alongside other models ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))

**Use: Obsidian setup**
- u/BehindUAll: recommends GPT-5.4 for Obsidian knowledge base work over Gemma4 ([reddit](https://reddit.com/r/hermesagent/comments/1tfak3b/creating_obsidian_knowledge_base/))

---

## GPT-4o

**Use: Legacy / comparison baseline**
- u/punkyrockypocky: "Different models may be more or less efficient with tokens for a given task so this isn't quite a 1:1 comparison" — referencing GPT-4o as baseline for cost analysis alongside other models ([reddit](https://reddit.com/r/hermesagent/comments/1tf3f2f/my_estimated_tokens_cost_saving_in_a_month/))

---

## Gemini 2.5

**Use: Setup and configuration assistance**
- u/TexBluBoy: "I used a combination of Gemini Pro & Gemini CLI for setting up my systems" ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))
- u/Affectionate-Permit9: "I asked Gemini pro what to ask Hermes to set it up and it's working great for 2 weeks" ([reddit](https://reddit.com/r/hermesagent/comments/1tfak3b/creating_obsidian_knowledge_base/))

**Use: Free model rotation**
- u/Little-Tea7664: "I usually rotate between free models and whatever seems to be getting good reviews" including Gemini ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))
- u/Hugo310 just set up Hermes with Gemini Flash via OpenRouter ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))

---

## Claude Sonnet / Opus

**Use: Premium coding and complex reasoning**
Claude is still the gold standard for quality, but cost keeps most users on alternatives. Mentioned primarily as the benchmark other models are compared against.

- u/_clickfix_: "With 128GB you can run the full GPT-OSS-120B model, which is as good as Claude Sonnet" ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))
- u/Colosteve2000: "The reason I say Sonnet or Opus is they are the only ones that do good at not losing context" ([reddit](https://reddit.com/r/hermesagent/comments/1tf3f2f/my_estimated_tokens_cost_saving_in_a_month/))
- u/Immediate_Let_4946: "Just the darn gpt5.5 and Claude Sonnet are superior" to Minimax ([reddit](https://reddit.com/r/hermesagent/comments/1tf3f2f/my_estimated_tokens_cost_saving_in_a_month/))
- u/Thomas-Lore notes Kimi "is close to Sonnet but only the largest models" — positioning Sonnet as the quality ceiling ([reddit](https://reddit.com/r/hermesagent/comments/1tegogu/running_hermes_with_local_models/))

---

## Grok (SuperGrok)

**Use: Subscription integration**
Controversial. Some users praise the integration, others find the models inferior to alternatives.

- Thread: "Hermes + SuperGrok is a beautiful marriage" — but u/hometechgeek suspects "this is another grok bot" ([reddit](https://reddit.com/r/hermesagent/comments/1tfrv4k/hermes_supergrok_is_a_beautiful_marriage/))
- u/EyeSuper7444: "I tried switching to Grok 4.3 for tool use... it wouldn't be smart enough" ([reddit](https://reddit.com/r/hermesagent/comments/1tfrv4k/hermes_supergrok_is_a_beautiful_marriage/))
- u/Delicious_Ease2595: "I tried Grok models in OpenClaw and I did not find them good enough compared to alternatives" ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))
- Thread: SuperGrok subscription now available on Hermes Agent ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))
- u/HobokenChickens uses Grok in a multi-model routing setup ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/Mighty_Buddha mentions Grok alongside other providers for privacy-focused users ([reddit](https://reddit.com/r/hermesagent/comments/1teqzjg/questionconcern_are_you_using_nous_or_other/))

---

## Mistral / Codestral

**Use: Agent configuration and specialized tasks**
- u/hoochiesan: "I've been doing this too!, 1 of 5 of my agents in telegram have this config. I have Mistral for..." — multi-agent setup with model specialization ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))
- u/wtfzambo had problems with "providers that had GLM" and switched approaches — Mistral mentioned as alternative in provider rotation ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))

---

# PROVIDERS & ROUTING

Not a model, but how you access them. Community strategy for getting the most out of your budget.

---

## OpenRouter

**Use: Multi-model access and free tier exploitation**
OpenRouter is the dominant provider platform. Users exploit free models, rotate based on availability, and manage credit carefully.

- Thread: "PSA for OpenRouter users" — 59 points, 16 comments. Key advice on caching and routing. ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))
- u/MrFretless5: "I've been using OpenRouter free models, with a $10 credit, and has been stable so far" ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- u/GreeneryCA: "Every morning test the free Openrouter options and chg to the best" ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))
- u/Hugo310 uses OpenRouter's Pareto router for automatic model selection ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/8bit64k: "I'm using OpenRouter and I've been very happy" ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- u/Sanky1120 had $11 in OpenRouter credits to work with ([reddit](https://reddit.com/r/hermesagent/comments/1te4qqt/psa_for_openrouter_users/))

**Use: Privacy-focused routing**
- u/Mighty_Buddha recommends venice.ai as alternative for privacy concerns ([reddit](https://reddit.com/r/hermesagent/comments/1teqzjg/questionconcern_are_you_using_nous_or_other/))

---

# MODEL SELECTION DISCUSSIONS

These threads cover model selection strategy rather than specific models:

- **"Battle of the $20 (or cheaper) providers"** — 119pts, 85 comments. The definitive cost-vs-quality thread. ([reddit](https://reddit.com/r/hermesagent/comments/1tewdky/battle_of_the_20_or_cheaper_providers/))
- **"Which model do you use with Hermes to balance token usage and reasoning quality?"** — 10pts, 18 comments. ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- **"Advice on model"** — 3pts, 40 comments. Newcomers asking what to run. ([reddit](https://reddit.com/r/hermesagent/comments/1tg1pal/which_model_do_you_use_with_hermes_to_balance/))
- **"My estimated tokens cost saving in a month. Need critiques."** — 3pts, 25 comments. Cost analysis breakdown. ([reddit](https://reddit.com/r/hermesagent/comments/1tf3f2f/my_estimated_tokens_cost_saving_in_a_month/))
- **"50k+ tokens spent on every single prompt... why?"** — 9pts, 17 comments. Token usage optimization. ([reddit](https://reddit.com/r/hermesagent/comments/1ted4a9/50k_tokens_spent_on_every_single_prompt/))
- **"Model Selection: Cold Outbound Email with Hermes"** — 2pts, 8 comments. Choosing models for writing tasks. ([reddit](https://reddit.com/r/hermesagent/comments/1tfm0qd/model_selection_cold_outbound_email_with/))

---

# KNOWLEDGE TABLES

## Local Models

| Model | Hardware Requirements | Use Cases | Community Sentiment | Cost | Notable Mentions |
|-------|----------------------|-----------|-------------------|------|-----------------|
| **Qwen 3.6 27B** | 16GB+ VRAM / 32GB RAM | Primary local, knowledge base, Obsidian | Very positive — "fantastic" | Free (local) | u/mrgreatheart, u/fuchelio, u/Britbong1492 |
| **Qwen 3.6 35B-a3b** | 8GB+ VRAM | Budget local, 8GB GPU friendly | Positive — "pretty adequate" | Free (local) | u/Thickdickmick87, u/Express_Nebula_6128 |
| **Qwen 3.5 122B** | 64GB+ RAM / multi-GPU | Large local, complex tasks | Mixed — struggles with some tasks | Free (local) | u/JBManos |
| **Gemma 4** | 32GB+ RAM | Creative/SillyTavern only | Negative for Hermes agent use | Free (local) | u/BehindUAll, u/PSyCHoHaMSTeRza |
| **Llama 4 Maverick** | 128GB+ RAM | Large local experiments | Positive for capable hardware | Free (local) | u/ButterflyEconomist, u/Rootshot |
| **GLM 5.1** | 16GB+ VRAM | Sweet-spot balanced tasks | Positive — "nice sweet spot" | Free (local) | u/itssethc, u/TralfamadorianNode |
| **Phi-4 Mini** | 4GB+ VRAM | Auxiliary helper model | Positive as helper | Free (local) | Referenced in profile configs |

## Cloud / API Models

| Model | Provider | Use Cases | Community Sentiment | Cost | Notable Mentions |
|-------|----------|-----------|-------------------|------|-----------------|
| **DeepSeek R1/V4** | DeepSeek native, OpenRouter | Daily driver, budget coding, flash tasks | Positive — best value for money | ~$1/day heavy use; free via Nous | u/mixxoh, u/torrso, u/SelectionCalm70 |
| **Minimax** | Minimax API, HuggingFace | Token-plan unlimited, budget coding | Mixed — cheap but "quite dumb lately" | $20-40/mo token plan | u/vandalieu_zakkart, u/yayita2500 |
| **MiMo V2 Pro** | Xiaomi API | Token-plan alternative, coding | Positive — better than Minimax per users | $16/mo 200M token plan | u/Ok_Firefighter3363, u/kawasaki500 |
| **Kimi K2.6** | Moonshot, OpenCode Go | Software dev, writing, Sonnet alternative | Positive — "close to Sonnet" | Via OpenCode Go sub | u/8bit64k, u/RawFreakCalm |
| **GPT-5.4** | OpenAI subscription | Coding, Obsidian setup, complex tasks | Positive but expensive | $20/mo subscription | u/dalemugford, u/BehindUAll |
| **GPT-4o** | OpenAI | Legacy comparison baseline | Neutral — older generation | — | u/punkyrockypocky |
| **Gemini 2.5** | Google, OpenRouter | Setup assistance, free rotation | Positive for setup tasks | Free tier available | u/TexBluBoy, u/Affectionate-Permit9 |
| **Claude Sonnet** | Anthropic, OpenRouter | Premium coding, quality benchmark | Gold standard but expensive | Premium pricing | u/_clickfix_, u/Colosteve2000 |
| **Claude Opus** | Anthropic direct | Complex reasoning, quality ceiling | Best quality, highest cost | Highest tier | u/Colosteve2000 |
| **Grok** | xAI, SuperGrok sub | General use, subscription integration | Mixed — "not good enough" vs alternatives | SuperGrok subscription | u/Delicious_Ease2595, u/EyeSuper7444 |
| **Mistral** | Mistral API | Multi-agent config, specialized tasks | Niche use, limited mentions | API pricing | u/hoochiesan |

---

# TOP CONTRIBUTORS

| User | Key Contributions |
|------|-----------------|
| u/mrgreatheart | Qwen 3.6 27B daily driver setup, uncensored variant discovery |
| u/torrso | OpenRouter PSA — DeepSeek caching advice |
| u/8bit64k | Kimi K2.6 as primary for software dev |
| u/vandalieu_zakkart | Minimax token plan honest review |
| u/Colosteve2000 | Multi-model cost analysis, Sonnet/Opus quality comparison |
| u/TralfamadorianNode | Dappnode Next subscription + local Qwen/GLM setup |
| u/TexBluBoy | GMKtec EVO-X2 hardware + Gemini for setup |
| u/BehindUAll | Gemma4 criticism, GPT-5.4 recommendation for Obsidian |
| u/Britbong1492 | 95% local Qwen routing system with cloud fallback |
| u/Immediate_Let_4946 | Minimax vs GPT-5.4 mini comparison |
| u/Ok_Firefighter3363 | MiMo V2 Pro token plan discovery |
| u/dalemugford | OpenAI Codex sub + o3 reasoning combo |
| u/Affectionate-Permit9 | Gemini Pro for Hermes self-configuration |
| u/fuchelio | Qwen 3.6 27B full precision for knowledge base |
| u/LouVillain | Minimax-M2.1 via HuggingFace, frontier provider experience |
| u/wtfzambo | Kimi vs Gemma4 comparison, GLM troubleshooting |
| u/RawFreakCalm | Kimi writing quality vs GPT assessment |
| u/yayita2500 | Minimax token plan + minimax 2.5 readiness |
| u/itssethc | GLM 5.1 sweet spot recommendation |
| u/hoochiesan | Multi-agent Telegram config with Mistral |

---

*Sources: 32 posts from r/hermesagent (Apr 30 – May 17, 2026). All user quotes are from public Reddit threads. Engagement scores reflect community consensus at time of collection.*