The method and what changed
I retrieved the public models API: 461 catalogue entries, including 13 models with zero text-token prices and one free router. Twelve models have the :free suffix; stealth/crouton-alpha is a temporarily zero-priced preview. openrouter/free selects from free models: it is a router, not another model.
The filter checks zero pricing.prompt and pricing.completion, then modalities and pricing descriptions. Correction to the July list: both Lyria entries are excluded because music clips are billed separately despite those zero fields. Qwen3 Coder, GPT OSS, Llama 3.3 and several other older variants no longer appear in today’s free selection. This does not mean their paid versions have disappeared.
This is a catalogue check, without a fresh inference test of every model. Capabilities below come from API listings, not a performance ranking. Read the dated snapshot and criteria.
Models to try by task
I would start with these candidates by task, then measure quality, speed and reliability on my own workload.
| Use case | Model | What the catalogue says |
|---|---|---|
| Coding and tools | recursive/flint-v1.0:free | Open-weight model aimed at coding agents; tool support listed. |
| Specialized coding | cohere/north-mini-code:free | Designed for agentic coding, with a 256,000-token context. |
| Long context | nvidia/nemotron-3.5-lightning:free | Advertised 1M-token context, text input and tool calling. |
| Vision and documents | nvidia/nemotron-3.5-super-vl:free | Text, image and video input, with an advertised 1M context. |
| Compact text model | liquid/lfm-3-1.7b:free | 1.7B parameters according to the listing; tools and structured output. |
| Multimodal | google/gemma-4-26b-a4b-it:free | Text, image and video inputs listed; text output. |
| Automatic selection | openrouter/free | Selects a free model compatible with the request. |
These models are candidates for workflows in DeepSeek Harness, with its plugin architecture and MCP connections, through a compatible OpenRouter provider. This is the kind of environment where I would compare models for research, extraction or coding. Check tool support on the actual endpoint: a large context alone does not make a good agent. These OpenRouter models were not run in my Harness comparison, which used DeepSeek Flash.
Free access list as of November 9
Exact IDs, sorted by decreasing context size. Context is the catalogue value, not a quality guarantee or an output limit.
| Model / router | Context tokens | Input → output |
|---|---|---|
stealth/crouton-alpha | 1,048,576 | text+image+file+audio+video → text |
nvidia/nemotron-3-ultra-550b-a55b:free | 1,000,000 | text → text |
nvidia/nemotron-3.5-lightning:free | 1,000,000 | text → text |
nvidia/nemotron-3.5-super-vl:free | 1,000,000 | text+image+video → text |
dots-studio/dots-3-note-preview:free | 512,000 | text+image → text |
google/gemma-4-26b-a4b-it:free | 262,144 | text+image+video → text |
google/gemma-4-31b-it:free | 262,144 | text+image+video → text |
liquid/lfm-3-1.7b:free | 262,144 | text → text |
recursive/flint-v1.0:free | 262,144 | text → text |
xingchen-agi/xing4.0-29b-a4b:free | 262,144 | text → text |
cohere/north-mini-code:free | 256,000 | text → text |
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free | 256,000 | text+image+audio+video → text |
liquid/lfm-2.5-2.6b:free | 65,536 | text → text |
openrouter/free | 200,000 | text+image → text |
History: the July 2026 list
Many models have left the free selection since the first version of this article. It included Qwen3 Coder, both GPT OSS models, Llama 3.3, Hermes 3 and Poolside models. Their former free IDs do not appear in the November snapshot, so keeping an unchanged workflow cannot depend on them.
To preserve a record, the 25 entries published in July remain below. Comparing like with like, July contained 22 :free variants: five remain, 17 are absent and seven new variants have arrived. That leaves 12 :free variants on November 9. This is substantial turnover, not just a smaller total.
Expand the original table — 25 entries on July 1, 2026
An archive, not a current availability list. The original table retains its published context and modality values. It included the openrouter/free router and two Lyria entries; Lyria has since been excluded from the free selection because music generation is billed separately. Those 25 rows were therefore not 25 :free variants.
| Model | Context | Modality |
|---|---|---|
google/lyria-3-clip-preview | 1M | text+image → text+audio |
google/lyria-3-pro-preview | 1M | text+image → text+audio |
qwen/qwen3-coder:free | 1M | text → text |
nvidia/nemotron-3-super-120b-a12b:free | 1M | text → text |
nvidia/nemotron-3-ultra-550b-a55b:free | 1M | text → text |
google/gemma-4-26b-a4b-it:free | 262K | text+image+video → text |
google/gemma-4-31b-it:free | 262K | text+image+video → text |
poolside/laguna-m.1:free | 262K | text → text |
poolside/laguna-xs.2:free | 262K | text → text |
qwen/qwen3-next-80b-a3b-instruct:free | 262K | text → text |
cohere/north-mini-code:free | 256K | text → text |
nvidia/nemotron-3-nano-30b-a3b:free | 256K | text → text |
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free | 256K | text+image+audio+video → text |
openrouter/free | 200K | text+image → text |
meta-llama/llama-3.2-3b-instruct:free | 131K | text → text |
meta-llama/llama-3.3-70b-instruct:free | 131K | text → text |
nousresearch/hermes-3-llama-3.1-405b:free | 131K | text → text |
openai/gpt-oss-120b:free | 131K | text → text |
openai/gpt-oss-20b:free | 131K | text → text |
nvidia/nemotron-3.5-content-safety:free | 128K | text+image → text |
nvidia/nemotron-nano-12b-v2-vl:free | 128K | text+image+video → text |
nvidia/nemotron-nano-9b-v2:free | 128K | text → text |
cognitivecomputations/dolphin-mistral-24b-venice-edition:free | 32K | text → text |
liquid/lfm-2.5-1.2b-instruct:free | 32K | text → text |
liquid/lfm-2.5-1.2b-thinking:free | 32K | text → text |
Read the archived July snapshot. Former IDs remain as historical references; free access is not guaranteed.
The limits of free access
- Quotas and availability. Free models have account and provider limits. The documentation recommends checking
free_model_daily_requeststhroughGET /api/v1/key. Provider capacity can still cause 429 errors. - Previews. Crouton Alpha may change price or disappear; zero pricing is not a permanent promise.
- Fallbacks. A paid fallback is billed at its own rate. To keep model inference free, select only free models or use the free router alone. Plugins and external services may have separate charges.
- Data. Check provider retention and terms before sending sensitive data. OpenRouter alone does not guarantee privacy.
Keep a model you like running locally
Like a model you tried on OpenRouter? Check whether its weights are available. You may then be able to download it from Hugging Face and run it locally with llama.cpp, Ollama or LM Studio, provided the architecture is supported.
Check the licence and choose a quantization compatible with the runtime and your RAM/VRAM. Leave memory for context too: very large models, including large MoEs, will not fit on a typical PC. Quantization may change quality, so rerun your task checks before migrating a workflow. See my experience with running open-weight models locally.
This reduces reliance on a free endpoint that may be limited or removed overnight. Keeping the weights, runtime version and configuration helps preserve a workflow you value, provided you have suitable hardware. A compatible local endpoint can then serve Harness too.
My advice: explore, test and keep an alternative
A lot has changed since the first list. The article showed 25 entries in July; today there are 12 :free variants, alongside a zero-priced preview and the free router. As the history explains, July’s total also included the router and both Lyria entries. Models leave and others arrive: free access is a moving target.
I would use these endpoints for discovery, practice, experiments, POCs and comparing what models can do. I would not make production depend on a free endpoint: quotas, capacity and removals can break a workflow overnight. For lasting use, plan suitable local hosting or a service whose commitments fit your needs; paying alone does not provide an SLA.
Data handling and sovereignty matter too. The prompts and content you submit pass through APIs and their hosting providers. Depending on the provider and your settings, they may be retained and potentially used for training. This is not universal: OpenRouter distinguishes its own policies from providers’ policies and lets you exclude providers that train on your data. Check retention and processing location too; I would use fictitious or non-sensitive data for these experiments.
It is still a useful way to try LLMs for free with other tools: configure a compatible endpoint in Hermes Desktop, explore plugins and MCP in DeepSeek Harness, or experiment with OpenCode and its model configuration. The OpenCode article covers an earlier experiment: choose an ID that is still available today. Check tool support, account for quotas and compare before committing to a model.
Sources and verification
Checked November 9, 2026:



