Krishna's Blog — AI, Technology & Software Engineering
← Back to Home

September 29, 2026 • AI Research

Free & Open-Weight AI Models: Complete Late-2026 Guide

The open-weight AI landscape has changed dramatically since February. New model generations have shipped from every major family, licensing has shifted (notably Gemma going Apache 2.0), OpenAI released its first open-weight LLMs in years, and entirely new families entered the field. This is a full refresh of the February 2026 guide — updated models, corrected licenses, and new recommendations.

A note on terminology: most models here are open-weight, not open-source. You get downloadable weights you can run, fine-tune, and build on — but not the training data or full training pipeline. Licenses vary more than marketing suggests; the table below gives the real terms per family.

What changed since February 2026

  • Llama 4 (Scout/Maverick, MoE) replaced Llama 3.x as Meta's current line
  • Gemma 4 moved to Apache 2.0 — a major licensing improvement over Gemma 3's custom terms
  • DeepSeek V4 / V4.1-Flash pushed efficiency further: 4x less KV-cache memory via a new encoder-decoder architecture
  • Qwen3 expanded into a full ladder: dense, MoE, coder, vision, and omni variants (plus a 397B Qwen3.5 MoE)
  • OpenAI released gpt-oss (20B/120B MoE, Apache 2.0) — its first open-weight language models since GPT-2
  • New families: Microsoft's Phi (small-model specialist, MIT), Moonshot's Kimi K3 (2.8T-parameter MoE), Z.ai's GLM-4.7 (coding-focused, MIT)

Quick Comparison Table

Family Provider Sizes Best For License
Qwen3 Alibaba 0.6B – 235B (dense + MoE); 397B in Qwen3.5 Coding, multilingual, full ladder Apache 2.0
Llama 4 Meta Scout 109B / Maverick ~400B (MoE, 17B active) General purpose, largest ecosystem Custom Community License (gated)
DeepSeek V4 DeepSeek 284B – 1.6T MoE (13B – 49B active) Reasoning, agentic coding, efficiency MIT
Mistral Mistral AI 3B – 675B (dense + MoE) Efficiency, EU hosting, enterprise Apache 2.0 (Small/Large); Modified MIT w/ monthly revenue cap (Medium)
Gemma 4 Google E2B, E4B, 12B, 26B MoE, 31B Lightweight, mobile/edge, multimodal Apache 2.0 (+ Prohibited Use Policy)
gpt-oss OpenAI 20B / 120B MoE OpenAI-compatible open weights Apache 2.0
Phi Microsoft 1.3B – 15B Small models, edge, reasoning MIT
Kimi K3 Moonshot AI 2.8T MoE (104B active) Long context (1M), agentic tasks Custom Kimi K3 License
GLM-4.7 Z.ai Varies Coding, multilingual coding MIT

1. Qwen3 (Alibaba)

Overview

Qwen is the broadest open-weight family: the original Qwen3 series spans 0.6B–235B across dense and MoE architectures, with dedicated coder, vision-language, and omni variants. Qwen3-Coder is officially billed as the Qwen team's "most agentic code model to date." The newer Qwen3.5 series adds a 397B MoE flagship.

Key Specifications

  • Sizes: 0.6B to 235B (dense + MoE); Qwen3.5-397B-A17B (397B total / 17B active) is the largest
  • Variants: Qwen3 (general), Qwen3-Coder, Qwen3-VL, Qwen3-Omni
  • Strengths: Agentic coding (Qwen3-Coder-480B-A35B-Instruct), 119 languages and dialects, 1M-token context via Qwen3-Coder extrapolation
  • License: Apache 2.0 — all Qwen3 models, per the official technical report

Download Links

2. Llama 4 (Meta)

Overview

Meta's current generation: Scout (109B total) and Maverick (~400B total), both MoE with 17B active parameters. Llama remains the default choice for the largest fine-tuning ecosystem — but the license carries restrictions that matter at commercial scale: gated downloads, a >700M monthly-active-user threshold requiring a separate Meta license, an Acceptable Use Policy, "Built with Llama" attribution, and EU restrictions on the multimodal models.

Key Specifications

  • Models: Llama 4 Scout (109B total / 17B active), Llama 4 Maverick (~400B total / 17B active), MoE
  • Strengths: Largest community, most fine-tunes and integrations, natively multimodal
  • License: Custom Llama 4 Community License — gated downloads, usage thresholds apply. Read it before commercial deployment.

Download Links

3. DeepSeek (V4 / R1)

Overview

DeepSeek continues to set the efficiency bar. V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B active) ship with 1M-token context and thinking modes; V4.1-Flash's Causal Encoder-Decoder architecture cut KV-cache memory ~4x versus V4-Flash (890 bytes/token). R1 remains the reference open reasoning model (671B / 37B active). All recent releases are MIT-licensed — the safest default for commercial use.

Key Specifications

  • DeepSeek-V4-Pro / V4-Flash: Flagship MoE, 1M context, agentic coding SOTA among open models
  • DeepSeek-V4.1-Flash: Efficiency-optimized — ~4x less KV-cache memory than V4-Flash
  • DeepSeek-R1: Advanced reasoning, math, logic — comparable to OpenAI's o1-class at release
  • Strengths: Reasoning quality near the closed frontier, cost efficiency, permissive MIT terms
  • License: MIT across recent releases (R1, V4, V4.1). The older V3 line used a custom DeepSeek license.

Download Links

4. Mistral AI

Overview

Mistral's lineup splits by license: Small and Large lines are Apache 2.0, while Medium ships under a Modified MIT license with a monthly revenue cap ($20M consolidated revenue in the preceding month). Still the efficiency-first European option, and the natural pick for EU data-residency requirements — La Plateforme hosts data in the EU by default. Codestral's 22B weights are non-commercial (MNLP-0.1 license).

Key Specifications

  • Sizes: 3B (Ministral/Voxtral edge models) to 675B (Mistral Large 3, 41B active MoE)
  • Models: Mistral Small 4 (119B MoE, Apache 2.0), Mistral Medium 3.5 (128B dense, Modified MIT), Mixtral MoE variants, Codestral
  • Strengths: Performance per parameter, 77.6% SWE-bench Verified (Medium 3.5), EU hosting
  • License: Apache 2.0 (Small 4 / Large 3); Modified MIT with monthly revenue cap (Medium 3.5) — verify which applies to your use

Download Links

5. Gemma 4 (Google)

Overview

The biggest licensing story of the year in open weights: Gemma 4 moved to Apache 2.0, replacing Gemma 3's custom terms (note: Apache 2.0 is paired with a separate Prohibited Use Policy — not a bare license). The lineup is now five sizes, including an MoE tier that activates only 3.8B parameters. Same lightweight, multimodal, multilingual strengths — 140+ languages — now with far less legal friction.

Key Specifications

  • Sizes: E2B, E4B, 12B Unified, 26B MoE (A4B — 3.8B active), 31B Dense
  • Variants: Pre-trained (base) and instruction-tuned; ShieldGemma 2 is a companion safety model, not a Gemma 4 variant
  • Strengths: Runs on consumer hardware and mobile, natively multimodal (video, images; audio on E2B/E4B), 140+ languages, 128K–256K context
  • License: Apache 2.0 (Gemma 4) plus the Gemma Prohibited Use Policy; older Gemma 1–3 used custom Gemma Terms

Download Links

6. gpt-oss (OpenAI)

Overview

OpenAI's first open-weight language models since GPT-2: two MoE models under Apache 2.0, natively quantized in MXFP4 with 128K context. The 120B model reaches near-parity with o4-mini on reasoning benchmarks; the 20B matches or exceeds o3-mini. For teams standardized on OpenAI's API formats, these are the natural self-hosted fallback — compatible with the Responses API and the open-sourced Harmony prompt format.

Key Specifications

  • Sizes: gpt-oss-120b (117B total / 5.1B active), gpt-oss-20b (21B total / 3.6B active)
  • Strengths: OpenAI-compatible behavior, permissive license, exposed chain-of-thought, tool use
  • License: Apache 2.0

Download Links

7. Phi (Microsoft)

Overview

The small-model specialist: Phi models (1.3B–15B, MIT) trained on curated textbook-quality synthetic data, designed for memory/compute-constrained and latency-bound scenarios. A full reasoning line exists (Phi-4-reasoning, mini reasoning, and a 15B multimodal reasoning variant).

Key Specifications

  • Sizes: 1.3B to 15B
  • Strengths: Best-in-class tiny models, reasoning variants, runs on phones and laptops
  • License: MIT

Download Links

8. Kimi K3 (Moonshot AI) & GLM-4.7 (Z.ai)

Overview

Two newer entrants worth tracking. Moonshot's Kimi K3 is a 2.8-trillion-parameter MoE (896 experts, 104B activated) with 1M-token context and native multimodal input, positioned as an agentic long-horizon coding model — under a custom Kimi K3 License, not a permissive OSI license. Z.ai's GLM-4.7 is a coding-focused open model (73.8% SWE-bench Verified) with strong multilingual coding performance, released under MIT.

Download Links

How to Run These Models Locally

Tool Best For Link
Ollama Easiest setup, command line ollama.com
LM Studio GUI interface, beginners lmstudio.ai
vLLM High-throughput serving, production inference GitHub
llama.cpp Maximum performance, developers GitHub
Hugging Face Python integration, ML workflows huggingface.co

Rule of thumb: the smallest model that passes your eval wins. Don't reach for 70B when a well-chosen 8B clears the bar — it's cheaper, faster, and fits more hardware.

Choosing the Right Model

For General Use

Llama 4 (Maverick-class) or Qwen3 (235B-class) — best balance of capability and ecosystem. Llama if you want maximum community support; Qwen if you want Apache 2.0 certainty.

For Coding

Qwen3-Coder and DeepSeek variants are purpose-built for code generation, debugging, and agentic coding workflows. GLM-4.7 is a strong MIT-licensed alternative.

For Limited Hardware

Gemma 4 (E4B) or Phi (3–4B) — run on consumer GPUs, high-end CPUs, even mobile. Now both under permissive licenses.

For Long Documents

Qwen3-Coder (1M via extrapolation), DeepSeek V4 (1M), or Kimi K3 (1M) — built for entire codebases and book-length context.

For Math and Reasoning

DeepSeek-R1 — still the open reference for step-by-step reasoning and proofs.

For Commercial Products

Prefer MIT / Apache 2.0 families (Qwen, DeepSeek, Gemma 4, gpt-oss, Phi, GLM-4.7, Mistral Small/Large). Treat Llama's Community License, Mistral Medium's revenue-capped terms, and Kimi K3's custom license as requiring legal review.

Important Considerations

  • Hardware: 70B+ models typically need 48GB+ VRAM or significant RAM for CPU inference; quantized 7–8B models run on a single consumer GPU
  • Quantization: GGUF/AWQ quants run on far less hardware with minimal quality loss — always check for quantized releases before buying GPUs
  • Licenses: "open" is not one thing. Apache 2.0/MIT are safe defaults; custom community licenses may restrict commercial scale or require gating
  • Freshness: Open-weight releases now ship monthly. Pin versions in production; re-evaluate quarterly

Last updated: September 29, 2026. Model specifications and availability change frequently — verify against the model card before deploying.

References

All model specifications, license terms, and version details above were verified against official vendor sources on September 29, 2026:

Comments & Reactions