Text models 314
We track 314 open text models from 35 labs, 87 of them released in the 12 months to 7 Nov 2026. The newest is Qwen3.8-27B-DSpine-b16 (7 Nov 2026) and the most downloaded is Qwen3-0.6B, with 19,141,315 downloads in 30 days. 172 fit a 16 GB GPU at 4-bit, and 43 have a hosted API.
llama.cpp, vLLM and MLX provide different ways to run text-generation models. Local use favours easy installation and memory fit; a shared endpoint also needs to handle concurrent requests. The model architecture, quantisation and runtime must be compatible before a published benchmark is useful.
What the category covers
Instruction-following chat models are the obvious case, but the same architecture class includes base models meant for further training, reasoning variants tuned to spend more tokens before answering, and small models in the sub-billion range aimed at phones and edge boxes. Mixture-of-experts designs sit in the same category: they change which parameters activate per token, not what the model consumes or emits.
How they are measured
EleutherAI’s Language Model Evaluation Harness provides a shared framework for evaluating generative language models across academic tasks. Its repository documents task definitions, model integrations and configuration controls. A common tool helps researchers reproduce an evaluation, but it does not make results from different settings interchangeable.
That matters more than any individual score. Benchmark numbers in a model card are only comparable when the harness, the few-shot count and the prompt template match, and the harness exists precisely so those conditions can be pinned down. When a lab reports a number without naming the evaluation setup, treat it as a claim rather than a measurement.
Scores also decay in usefulness as they leak into training data. Contamination-aware benchmarks address this by collecting fresh problems over time rather than publishing a fixed set, an approach LiveCodeBench applies on the coding side.
What serves them
vLLM supports GPU serving through CUDA, ROCm and Intel XPU paths, with Apple GPU support through the separate vLLM-Metal plugin. Its installation overview also lists CPU targets and other hardware plugins. Each backend has its own installation and model-support limits.
llama.cpp takes the other approach, targeting local and cloud inference across a range of hardware. Its feature list includes integer quantisation and CPU plus GPU hybrid inference to partially accelerate models larger than total video memory. Getting a model running is one command: llama cli -hf with a repository name, or llama serve -hf for an OpenAI-compatible server.
On Apple silicon, mlx-lm is the native path, described by its README as a Python package for generating text and fine-tuning language models, with Hugging Face Hub integration, quantisation and LoRA. Our MLX page covers the commands.
Ollama wraps llama.cpp-style local running in a simpler interface, covered on our Ollama page, and the GGUF format those tools consume is explained on our GGUF page.
The hardware reality
Weights have to fit somewhere, and memory limits you before compute does. A model that does not fit in video memory either spills to system memory, which llama.cpp supports through hybrid inference at a speed cost, or does not run. Quantisation is the usual lever, trading some accuracy for a smaller footprint.
Apple silicon behaves differently because of unified memory: MLX arrays live in shared memory, so a Mac with more RAM can load larger models without a separate video memory budget. Memory bandwidth and runtime overhead still affect speed.
For serving rather than single-user chat, memory also goes to the key-value cache. vLLM’s optimization guide says cache space can run short when requests are batched, in which case it preempts and recomputes requests, and its memory guide lists a lower context length and a smaller maximum batch size among the ways to reduce memory use. A model that runs as a local chat assistant can need different settings as a shared endpoint on the same card.
Licence patterns
Model licences differ. For the conditions attached to permissive releases, see our pages on Apache 2.0, MIT and the comparison. Custom community licences add conditions such as naming requirements or use restrictions, which is the pattern on our Llama Community Licence and Gemma terms pages. A third group uses MIT with additions, covered on our modified MIT page.
Check the licence file in the repository rather than the label on the model card, because the two disagree often enough to matter.
This summarises licence text. It is not legal advice.
FAQ
Which runtime should I start with?
For a local machine, llama.cpp or Ollama, because quantised weights and hybrid inference make modest hardware workable. For a GPU server with concurrent users, vLLM. On a Mac, MLX.
Are benchmark scores in model cards comparable?
Only when the evaluation setup matches. The Language Model Evaluation Harness exists to make runs reproducible, and cards that do not state the harness, task version and shot count are not comparable to each other.
Do mixture-of-experts models need less memory?
Fewer active parameters can reduce computation per token, but that does not reduce the full checkpoint to its active-parameter size. Memory use also depends on quantisation, expert placement and offloading.
Can one model handle text and images?
Some do, and those are a separate category on this site. See our vision-language page for how they are measured and served.
Pick a view
All modelsNewest
| Model | Released | Type | Size | Memory at 4-bit | 16 GB GPU | Downloads, 30 days | Cheapest API, blended / 1M |
|---|---|---|---|---|---|---|---|
| 7 Nov 2026 | LLM | 2B | 1.5 GB | Fits | 4,332 | No tracked API | |
| 26 Oct 2026 | LLM | 11.2B | 8 GB | Fits | 189 | No tracked API | |
| 21 Oct 2026 | LLM | 286B | 189.7 GB | Too big | 3,372 | No tracked API | |
| 21 Oct 2026 | LLM | 286B | 189.7 GB | Too big | 1,157 | No tracked API | |
| 17 Oct 2026 | LLM | 7.8B | 5.5 GB | Fits | 170 | No tracked API | |
| 15 Oct 2026 | LLM | 30.1B | 20.7 GB | Too big | 7,755 | No tracked API | |
| 14 Oct 2026 | LLM | 8.8B | 7.3 GB | Fits | 6,349 | No tracked API | |
| 14 Oct 2026 | LLM | 30B | 22.2 GB | Too big | 5,081 | No tracked API | |
| 2 Oct 2026 | LLM | 78.1B | 51.7 GB | Too big | 2,894 | No tracked API | |
| 27 Sep 2026 | LLM | 311B | 206.4 GB | Too big | 20,694 | No tracked API | |
| 27 Sep 2026 | LLM | 1.02T | 679.9 GB | Too big | 20,162 | No tracked API | |
| 21 Sep 2026 | LLM | 311B | 206.4 GB | Too big | 57,506 | $0.14 (₹13) |
Most downloaded
| Model | Released | Type | Size | Memory at 4-bit | 16 GB GPU | Downloads, 30 days | Cheapest API, blended / 1M |
|---|---|---|---|---|---|---|---|
| 27 Apr 2025 | LLM | 752M | 1.5 GB | Fits | 19,141,315 | No tracked API | |
| 27 Apr 2025 | LLM | 14.8B | 11.2 GB | Fits | 13,011,347 | $0.13 (₹12) | |
| 1 Dec 2025 | LLM | 685B | 468.1 GB | Too big | 8,442,070 | $0.234 (₹22) | |
| 16 Sep 2024 | LLM | 7.6B | 5.5 GB | Fits | 6,444,573 | No tracked API | |
| 27 Apr 2025 | LLM | 8.2B | 6.7 GB | Fits | 6,165,334 | No tracked API | |
| 4 Aug 2025 | LLM | 20.9B | 14 GB | Fits | 5,673,703 | $0.0225 (₹2.12) | |
| 17 Sep 2024 | LLM | 1.5B | 1.3 GB | Fits | 5,145,856 | No tracked API | |
| 16 Sep 2024 | LLM | 494M | 0.4 GB | Fits | 4,292,272 | No tracked API | |
| 18 Jul 2024 | LLM | 8B | 5.7 GB | Fits | 4,198,546 | No tracked API | |
| 27 Apr 2025 | LLM | 32.8B | 24 GB | Too big | 3,741,049 | No tracked API | |
| 10 Mar 2025 | LLM | 1000M | 0.7 GB | Fits | 3,406,623 | No tracked API | |
| 4 Aug 2025 | LLM | 117B | 77.4 GB | Too big | 2,433,479 | $0.045 (₹4.24) |
Smallest that fit 16 GB
| Model | Released | Type | Size | Memory at 4-bit | 16 GB GPU | Downloads, 30 days | Cheapest API, blended / 1M |
|---|---|---|---|---|---|---|---|
| 2 Dec 2025 | LLM | 0M | 1.2 GB | Fits | 543 | No tracked API | |
| 13 Oct 2024 | LLM | 9M | 0 GB | Fits | 5,899 | No tracked API | |
| 13 Oct 2025 | LLM | 32M | 0 GB | Fits | 255 | No tracked API | |
| 17 Apr 2024 | LLM | 128M | 0.1 GB | Fits | 8,685 | No tracked API | |
| 13 Oct 2025 | LLM | 186M | 0.1 GB | Fits | 152 | No tracked API | |
| 7 Feb 2024 | LLM | 248M | 0.3 GB | Fits | 1,308 | No tracked API | |
| 15 Apr 2024 | LLM | 248M | 0.3 GB | Fits | 200 | No tracked API | |
| 7 Feb 2024 | LLM | 248M | 0.3 GB | Fits | 136 | No tracked API | |
| 11 Apr 2024 | LLM | 248M | 0.3 GB | Fits | 129 | No tracked API | |
| 28 Jun 2025 | LLM | 361M | 0.4 GB | Fits | 8,875 | No tracked API | |
| 28 Jun 2025 | LLM | 361M | 0.4 GB | Fits | 2,113 | No tracked API | |
| 30 Mar 2026 | LLM | 438M | 0.5 GB | Fits | 2,840 | No tracked API |
Releases per month
87 in 12 months| Month | Total |
|---|---|
| Dec 2025 | 11 |
| Jan 2026 | 6 |
| Feb 2026 | 9 |
| Mar 2026 | 7 |
| Apr 2026 | 10 |
| May 2026 | 4 |
| Jun 2026 | 10 |
| Jul 2026 | 5 |
| Aug 2026 | 10 |
| Sep 2026 | 6 |
| Oct 2026 | 8 |
| Nov 2026 | 1 |
Cheapest hosted APIs
All API prices
gpt-oss-20b$0.0225 (₹2.12)
DeepSeek-V4-Flash$0.0314 (₹2.96)
gpt-oss-120b$0.045 (₹4.24)
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16$0.0488 (₹4.60)
c4ai-command-r7b-12-2024$0.0656 (₹6.19)
phi-4$0.07 (₹6.60)
Llama-3.2-1B-Instruct$0.0705 (₹6.65)
granite-4.2-8b$0.0825 (₹7.78)
Blended is 3 parts input to 1 part output, USD per million tokens, with INR at ₹94.29 per $1.