The open-weight OpenAI model that a real laptop can actually hold: a 14 GB download that Ollama says runs with as little as 16 GB of memory. The cheapest way to feel how an LLM behaves when nothing leaves your machine.
The big sibling gpt-oss:120b is a 65 GB download and wants an 80 GB GPU, so that one is cloud only. Be honest about what local buys you: privacy, offline work, zero cost and understanding — not frontier agent quality.
Free: pull it with Ollama and run it offline. Google Antigravity also lists gpt-oss-120b on its free Individual plan, so you can try the big one without hardware.
The open-weight Google family, and the most laptop-friendly one on this page: sizes on Ollama run from e2b at 7.2 GB through 12b at 7.6 GB to 31b at 20 GB, with vision, tool use and thinking.
Context is listed as 128k to 256k depending on size. Be careful with blog guides claiming the e2b size needs only about 3 GB of RAM — that conflicts with the 7.2 GB download, so test it on a real laptop before you promise anything to a room full of students.
Free to download and run via Ollama.
DeepSeek V4.1-Flash
DeepSeekSerious quality at roughly a tenth of Western flagship prices: 1M context, 384k max output, thinking and non-thinking modes, vision, tool calls and JSON output. It also speaks the Anthropic API format, so Anthropic-compatible clients can be pointed straight at it.
- 1000k context
- $0.3 / $1.2 in / out per 1M tokens
- open weights
The prices shown are peak. Off-peak is half: peak runs 01:00-04:00 and 06:00-10:00 UTC on weekdays, which in Kazakhstan (UTC+5) means 06:00-09:00 and 11:00-15:00 local — your evenings and weekends are the cheap window. And read the second half of open weights carefully: the model card describes a 552B backbone, so open does not mean it fits on your machine.
The weights are MIT-licensed on Hugging Face, so it is free if you own hardware for a 552B-parameter model. The hosted API is not free, but it is the cheapest serious API here.
The bigger DeepSeek model (V4-Pro-0813, listed at 1.7T parameters on Hugging Face) with the same 1M context and 384k max output, still cheaper than any Western flagship.
- 1000k context
- $1.32 / $3.96 in / out per 1M tokens
- open weights
No vision, unlike Flash. Same peak and off-peak split. It costs about four times what DeepSeek Flash does, so start with Flash and only move up if you can show that Flash is the thing failing.
The small open-weight model Meta did ship: 30B parameters under Apache 2.0, released 10 August 2026, with a 128k context. Proof that Meta ships nothing open any more is also wrong.
The download on Ollama is 18-19 GB. By the same rule of thumb that lets gpt-oss:20b (14 GB) run on 16 GB of memory, you want noticeably more than 16 GB here. 30B is not a frontier model — use it to learn how these things behave, not to ship a product.
Free to download and run via Ollama, where it already has over 200 thousand pulls.
The last open Llama, released 5 April 2025: 17B active parameters of 109B total, and a 10M-token context window that is still the largest number on this page by a factor of ten.
- 10000k context
- open weights
History, not a recommendation. Meta moved to closed weights in 2026 and there has been no Llama 5. Its sibling Maverick is 17B active of 400B total with 1M context. One more reason it matters to us: Llama-family tokenizers are the worst of the lot for Kazakh, about 4.73 tokens per word against 2.56 for Russian and 2.23 for Turkish.
Mistral Medium 3.5
MistralThe serious European open-weight option: a dense 128B model under a modified MIT licence, released 28 April 2026. If the requirement is that data never leaves your own servers, this is the tier that can actually do the work.
- 256k context
- $1.5 / $7.5 in / out per 1M tokens
- open weights
The 256k context is the smallest among the big labs — everyone else here sits at 1M, and that gap is felt immediately in agent work on a large repo. Self-hosting is also not a laptop story: the Ollama download is 80 GB.
The weights are downloadable under a modified MIT licence, so self-hosting is free of licence cost. The hosted API is paid.
Apache 2.0 and cheap. The permissive licence is the point: you can use it commercially with no strings, self-hosted or through the API, which is rare at this quality level.
- 256k context
- $0.15 / $0.6 in / out per 1M tokens
- open weights
Still the Mistral 256k ceiling. On the Mistral API batch processing is half price and cached input is 90% cheaper, which is the most aggressive caching discount in this catalogue.
Weights are Apache 2.0, so downloading and self-hosting costs nothing but hardware.
Laptop-sized and Apache 2.0, with the unusual property that input and output cost exactly the same — which makes it the easiest model in the catalogue to budget for.
- $0.15 / $0.15 in / out per 1M tokens
- open weights
The family is 3B at $0.10 / $0.10, 8B at $0.15 / $0.15 and 14B at $0.20 / $0.20. Use small models for one narrow job at a time. An agent loop needs 64k or more of context and dependable tool calling, and that is exactly where models this size start to wobble.
Weights are Apache 2.0 and the sizes are small enough for a normal laptop.
Qwen3.8 (open weights, 27B)
Alibaba (Qwen)The open-weight Qwen line is the most-pulled non-Western family on Ollama, and 27B is the size a strong laptop or a single decent GPU can actually hold.
The older Qwen3.6 (27B and 35B) still has 6.7M pulls against 2.3M for Qwen3.8, so the community default is one generation behind — worth checking which one your tutorial actually assumes. The Ollama guide for running Claude Code against a local model uses a Qwen model in its own example.
Free to download and run via Ollama.
The highest-scoring open-weight model on the independent Artificial Analysis Intelligence Index: 45 at max effort on v4.3.2, against 53 for Claude Fable 5.1 and GPT-6 Astra.
Two teaching points. First, the open-weight gap behind the frontier is visible but modest, and 66 of the 149 models the index evaluates are open-weight. Second, on Ollama this one exists only as a cloud model — open weights never meant it fits on your machine. Scores depend on the effort setting, so always name the setting when you quote one.
The second-strongest open-weight score on the Artificial Analysis index (44 at max effort), and available as a model choice inside Cursor alongside Claude, GPT and Gemini.
Cloud-scale like GLM-5.3: Ollama carries the Kimi line only as cloud models. The practical lesson from both is that the open-weight frontier and the runs-on-my-laptop frontier are two completely different lines, roughly 500 billion parameters apart.