Best Local AI Models 2026: Liquid LFM3 and Reward-Aligned Distillation

The best local AI model in November 2026 is the smallest one that passes your own evaluation, and for many narrow tasks that is now a model of about 2 billion parameters. Liquid AI released six such models on 27 October, and a research paper on 10 October reported a better way to distil large models into small ones. Both claims come with limits. Liquid ran its own benchmarks, and the paper tests maths and code on academic sets.

This post separates what has been published from what is inferred. We read Liquid's launch post, the LFM3-1.7B model card and its licence file, and the distillation paper's abstract and experiment section. Where a figure comes from press coverage only, we say so.

What is the best local AI model in 2026?

There is no single best, because "best" depends on your hardware, your task and your licence. A useful way to choose is three tiers, from smallest to largest.

• Around 2B parameters, for phones, laptops and edge devices. LFM3-1.7B is the newest option we verified. Its model card says it fits in about 1 GB at 4-bit precision.
• Around 27B, for a workstation or a single large GPU. Qwen3.8-27B is Apache-2.0 licensed according to its Hugging Face card, with a 262,144-token native context. We cover it in our guide to running Qwen 3.8 27B locally.
• Much larger open models, for a server. These need real GPU memory and are a hosting decision, not a local one.

We are not publishing a ranked list. A ranking needs one shared benchmark run on one machine, and the figures in vendor announcements come from different setups. Treat any "best local model" table, including ours, as a shortlist to test.

Licence matters as much as accuracy. The LFM Open License v1.0 defines a revenue threshold of 10 million US dollars in annual revenue. Commercial use by an entity at or above that threshold is not licensed under the agreement, and revenue of affiliated entities counts towards the total. The file also has a carve-out for qualifying non-profits, for non-commercial and research use. If your company or its parent is above the threshold, LFM3 is not a free choice. This is not legal advice: read the licence file and ask counsel.

What is Liquid LFM3?

LFM3 is a family of six open-weight models from Liquid AI, announced on 27 October 2026. All six are built on a 1.7-billion-parameter base called LFM3-1.7B-Base, according to the launch post and press coverage.

We could verify the text model on Liquid's own pages. The launch post we fetched was truncated, so the sizes of the other five models come from unite.ai's coverage of the model card, not from a page we read in full. Treat those rows as reported.

What Liquid says about LFM3-1.7B, all vendor-reported:

• It beats the earlier LFM2.5-2.6B "while using 35% fewer parameters".
• It scores 38.42 on WebArena-Verified, against 3.02 for LFM2.5-2.6B (reported via unite.ai and listed on the model card).
• It averages 49.37 across the benchmark table, ahead of Qwen3.5-9B at 47.24.
• Its KV cache for a full 128K context is 1.6 GB, against 4.3 GB for Qwen3.5-4B, per Liquid's post.

Liquid ran and published all of these benchmarks itself. Coverage notes no independent replication as of 27 October. We found none either.

The same pages are candid about weaknesses, and these are the numbers to read first. On GPQA Diamond the model scores 63.37 against 80.40 for Qwen3.5-9B. On SWE-bench Verified (OpenCode) it scores 43.40 against 51.20. On MCP-Atlas it scores 32.57 against 42.67. On RULER at 128K it scores 65.56 against 87.11. Liquid's post says larger models still win on agentic coding and knowledge. The model card says the model is "not recommended for agentic coding or complex multi-step reasoning due to its small size" and suggests adding retrieval for knowledge-heavy tasks.

What is reward-aligned distillation?

Reward-aligned distillation (RAD) weights token-level teacher supervision on student-generated responses by its predicted effect on the student's reward after an update. The paper is arXiv 2610.14572, submitted on 10 October 2026 by Yilan Chen and colleagues.

Some background. Distillation trains a small student model to imitate a larger teacher. Standard on-policy distillation gives the student dense, token-by-token guidance from the teacher, but it trains on every token, including tokens where the teacher's preference does not help the student do better on the task. Earlier methods pick tokens using proxies such as uncertainty or teacher-student disagreement. The authors say those proxies may not find the supervision that actually raises reward.

RAD instead weights each token by how much it is predicted to raise the student's reward after an update. Per the abstract, it does this by backpropagating through a temporary optimiser step, then adjusting token weights by projected ascent. The weights can boost, dampen or even flip individual gradient contributions.

The abstract reports these results, at the same training-response budget:

• On six competition-maths benchmarks, mean accuracy improves over the strongest baselines by 6.5 percentage points for Qwen3-1.7B (non-thinking) and 2.5 points for Qwen3-4B-Instruct-2507.
• On LiveCodeBench-v6, RAD reaches 54.69% and 58.67% pass@1 for Qwen3-1.7B (non-thinking) and OLMo3-7B-Instruct, exceeding the strongest baselines by 8.7 and 6.4 points.

Three cautions. The abstract does not name the teacher or the benchmarks, but the full paper does (section 5.1). The maths teacher is Qwen3-30B-A3B-Instruct-2507, and the six evaluation sets are AIME24, AIME25, AIME26, HMMT2025-Feb, HMMT2025-Nov and OlympiadBench. The paper reports that RAD takes 17 to 18% more training time than some baselines on Qwen3-1.7B maths, so it is not free. We read sections 3.2 and 5.1 only. The gains are measured against other training methods, not against a larger model. And it is a single academic result on small models. It is a reason to expect better small models, not evidence that your task will improve. For the mechanics and economics of distillation in general, see our distillation explainer.

Can a 2B model run an agent on a phone?

Partly, and the honest answer is "for narrow agents, with checks". Liquid designs LFM3-1.7B for "local agents, tool calling, structured outputs, and domain-specific fine-tuning", and it was post-trained for agent harnesses such as Hermes Agent, Pi and OpenClaw, per the model card. Its tool-use scores are reasonable: 71.39 on BFCL V4 and 80.36 on ToolSandbox, both vendor-reported.

Liquid publishes vendor-run benchmarks on an iPhone 17 Pro (Apple A19 Pro), alongside two other devices, an AMD Ryzen AI Max+ 395 and an Apple M5 Max. The model card specifies Q4_0 GGUF files in llama.cpp, a 32,768-token context, batch size one, 100 output tokens and GPU offload. It says LFM3-1.7B prefills at least 2.5 times faster and decodes at least 1.8 times faster than the compared models on all three devices, with the least memory. The exact figures are in a chart we could not read as text. These measurements show handset execution. They do not establish the reliability of a complete phone agent. Liquid's throughput tests used a single NVIDIA B300 GPU.

Our inference: a 2B model can handle single-step tool calls with a fixed schema, classification, extraction and short structured outputs. It is a poor fit for long chains of dependent steps, where a small error early compounds. The model's own card says as much. Our post on running a 27B model on a phone with 1-bit quantisation shows a related trap: PrismML reports category averages of 66.03 for the binary build and 80.00 for its full-precision reference on agentic and tool-calling benchmarks. Its maths average falls from 95.33 to 91.66, a 3.67-point decrease. Compression and small size hurt unevenly, so test the skill you need.

For production edge work, see our guide to on-device models with the Connect embedded SDK.

Distillation vs fine-tuning: which makes a better small model?

They solve different problems, and you often use both. Fine-tuning updates a pretrained model using additional training. Supervised fine-tuning uses labelled examples, and continued pre-training can use unlabelled text. Distillation supplies teacher training signals and can form part of that adaptation.

Teacher licence terms matter. Some model licences restrict using outputs to train another model, so check the teacher's terms before you build a pipeline on it. The related question of whether to fine-tune at all is in our post on when to fine-tune on your own data.

One new technique worth watching is weight editing by a frontier model, which MIT researchers describe in a research claim critics dispute. We explore it in our post on GPT-6 Astra and weight-space programming. It is research, not a method for production.

When is a small local model enough?

Use the model's published weaknesses and your own error costs to decide. This table is our judgement, not a benchmark.

The pattern is to start local and escalate on failure or low confidence. Local costs no per-token fee, but hardware and energy still cost money, as Liquid's own post notes.

How do you test a small model on your own task?

Build a test set before you compare models. Twenty to fifty real cases is a start, taken from your actual inputs, with the answer you would accept for each.

1. Write the pass rule first. Exact match, schema-valid, or a reviewer's judgement. Decide it before you see outputs.
2. Include the hard 20%. Edge cases, messy input, and cases where the right answer is "I do not know".
3. Run the candidates under the same settings. Same prompt, quantisation and context length. Liquid's own comparison used vLLM in BF16 for the competing models, which is not how you will run them on a laptop, so repeat it on your target setup.
4. Measure failures, not only the average. Count parse failures, refusals and wrong-but-confident answers.
5. Set a route-up threshold. If the small model fails the rule, or reports low confidence, send the case to a bigger one.

A model that scores 49.37 on someone else's average tells you little about your invoices, your tickets or your code.

What to watch next

Three signals will tell you whether the small-model trend is real for your work. First, independent replication of Liquid's benchmarks, which we did not find as of 9 November. Second, scrutiny of RAD's published full paper and any released implementation, including its teachers, evaluation datasets and training costs. Third, independent phone measurements covering sustained performance, memory use and battery cost. Until those arrive, hold the headline figures loosely and trust your own test set.

Where Swfte sits

Swfte Cortex is a governed AI desktop that runs on your laptop. Cortex knowledge bases run on the device, with retrieval over local files, folders and URLs, and local embeddings through Ollama give a fully local path. That pairs naturally with a small local model for knowledge-heavy work, since LFM3's own card recommends retrieval for knowledge-heavy tasks.

When a local model is not enough, Swfte Connect offers one API across many providers, with fallback to another provider in the chain when one is down or rate-limited. You can bring your own keys. That lets you keep the escalation path from the table above in one place.

We do not publish benchmark figures for any local model on Cortex, and this post does not claim any. Check Cortex's current model support before you plan around a specific one.

---

Related: Distillation explained, Run a 27B model on a phone, Edge AI with the Connect embedded SDK, GPT-6 Astra and weight-space programming.

Sources: Liquid AI, LFM3 launch post · Hugging Face, LFM3-1.7B model card · Hugging Face, LFM Open License v1.0 · Unite.AI, Liquid AI launches LFM3 · arXiv 2610.14572, Reward-Aligned Distillation · Hugging Face, PrismML Bonsai-27B 1-bit card · arXiv, RAD full paper · Hugging Face, Qwen3.8-27B · accessed 9 November 2026.

Method note: we read these pages on 9 November 2026. Liquid's launch post was truncated in our fetch, so the sizes of five of the six models are reported through Unite.AI and are not confirmed. Liquid's benchmarks are vendor-run and not independently replicated as far as we found. The RAD headline figures come from the arXiv abstract, and the teacher, datasets and cost details from sections 3.2 and 5.1 of the full paper. The task table is our judgement.