45 Open-Weight Models for Fine-Tuning: A 2026 Catalog Guide

How to read this list

Choosing a base model is about matching four constraints: the task your fine-tune has to solve, the inference hardware you'll run on, the training budget you have, and the licensing your deployment requires. This post organizes 45 catalog entries across 17 families by size class with those constraints visible. It does not assign a universal score or rank.

Models are listed with the catalog's minimum tier, any listed per-model surcharge, the license, and a concise selection note. Public checkout is paused, so inclusion is an intake candidate—not a promise of availability, feasibility, or acceptance.

Model availability and benchmark leaderboards change quickly. This guide was checked on August 25, 2026; vendor-reported scores are labeled, and the linked model card remains authoritative for architecture, license, modality, and context. A catalog entry is not proof that a fixed GPU configuration or quoted training method is feasible before intake.

This is a model-selection guide, not a framework-buying guide. If you already know the model and need OpenClaw, Hermes Agent, or a custom agent runtime configured around it, use the relevant project's documentation and open skill surfaces such as the Hermes skills documentation. Logos chooses and trains the model; managed business workflows route through Icarus.

Sub-10B — fast iteration, prototyping, edge deployment

The Starter tier's home. Many of these models can run on a consumer GPU after choosing a suitable precision or quantization, but memory needs vary by weights, context, cache, runtime, and adapter method. Training time is estimated at intake.

Qwen 3.5 2B (Starter, no surcharge, Apache 2.0)

The smallest Qwen 3.5 entry in the Logos catalog; the official family also includes a 0.8B release. Its smaller size makes it a candidate for fast prototyping, subject to task evaluation.

Qwen 3.5 4B (Starter, no surcharge, Apache 2.0)

Qwen 3.5 Small (March 2026). 4B with 256K context, 201 languages, thinking/non-thinking modes. Edge-ready efficient reasoning.

Gemma 4 4B E4B (Starter, no surcharge, Apache 2.0)

Google's E4B is the smallest Gemma 4 entry in the Logos catalog, not the smallest member of the complete family. It supports vision and audio with a 128K-token context window according to the official Gemma 4 model card. Consumer and edge feasibility depends on precision, context, runtime, and available memory.

Mistral 7B Instruct v0.3 (Starter, no surcharge, Apache 2.0)

A widely used 7B baseline with many public recipes. That ecosystem makes it a practical comparison point for style-transfer and classification work; it is not universally the “most fine-tuned” or safest model.

Qwen 2.5 7B Instruct (Starter, no surcharge, Apache 2.0)

An established prior-generation baseline with extensive public documentation and community recipes. Compare it with newer releases when recipe maturity matters.

Llama 3.1 8B Instruct (Starter, no surcharge, Llama 3.1 license)

A widely deployed 8B instruction baseline with broad runtime support. Its smaller artifact can shorten experiments relative to larger models, but delivery time is confirmed at intake. The Llama license includes an additional permission condition above 700 million monthly active users.

DeepSeek-R1-Distill-Llama-8B (Starter, no surcharge, MIT)

Reasoning-distilled from R1 onto a Llama-8B backbone. Evaluate its answer quality and latency on your task; hidden reasoning traces are not a persistent model artifact that Logos can promise to preserve.

Llama-3.1-Nemotron-Nano-8B (Starter, no surcharge, Llama 3.1 license)

NVIDIA's reasoning-tuned 8B Llama. Competitive with much larger models on agentic tasks. Pick this over vanilla Llama 3.1 8B when tool-calling and multi-step reasoning are the dominant workload.

Qwen3 8B (Starter, no surcharge, Apache 2.0)

Qwen 3 dense with thinking/non-thinking dual-mode reasoning. Apache 2.0 licensed alternative to Llama 3.1 8B when the Llama 700M MAU restriction is a concern.

Qwen 3.5 9B (Starter, no surcharge, Apache 2.0)

Qwen 3.5 Small (March 2026), with a 256K context window and multimodal and agent-oriented capabilities reported by the publisher. Evaluate its coding behavior and runtime support on the intended workload.

Hermes 3 Llama 3.1 8B (Starter, no surcharge, Llama 3.1 license)

Nous Research's fine-tune of Llama 3.1 8B — function-calling and agentic behavior pre-trained into the model, minimally filtered, highly steerable. Strong starting point for assistant-style products where you'd otherwise spend your fine-tune just teaching function-calling.

12B to 24B — mid-size dense models

Mistral Nemo 12B Instruct (Studio, no surcharge, Apache 2.0)

Balanced reasoning + long-context baseline. Good middle-of-the-road choice when you want more ceiling than 8B but don't need 32B-class compute.

Gemma 3 12B Instruct (Starter, no surcharge, Gemma license)

Google's 12B Gemma 3 instruction model. Evaluate its safety behavior and hardware footprint for the intended precision, runtime, and policy rather than treating “safe” as a model-wide guarantee.

Phi-4 14B (Studio, no surcharge, MIT)

Top-tier reasoning in a small footprint — punches well above its size class on math, logic, and code benchmarks. Pick Phi-4 when reasoning quality per inference dollar is the primary concern.

Qwen 2.5 14B Instruct (Starter, no surcharge, Apache 2.0)

Middle-class Qwen 2.5 — bridges the gap between 7B and 32B. Heavily documented recipes, strong general-purpose baseline. A Starter-tier 14B is rare; this is the most-cost-effective mid-size option.

Qwen3 14B (Starter, no surcharge, Apache 2.0)

Qwen 3 dense 14B — direct competitor to Phi-4 14B with Apache 2.0 licensing and dual-mode reasoning. Newer generation than Qwen 2.5 14B.

Mistral Small 24B Instruct (Studio, no surcharge, Apache 2.0)

Modern dense model sitting between Phi-4 14B and Qwen 32B. Good long-horizon choice for customer-facing agents — the Mistral family's tool-use behavior is predictable and the 24B size is cheap enough to run at scale.

27B to 36B — mid-frontier dense and MoE

Gemma 3 27B Instruct (Studio, no surcharge, Gemma license)

Frontier Gemma of the prior generation. Fills the 27B mid-class gap with Google's training.

Qwen 3.5 27B (Studio, no surcharge, Apache 2.0)

Current Qwen generation in the dense 27B class (released Feb 2026). Pick over Qwen 2.5 32B when you specifically want 3.5's updated training data and safety profile.

Qwen 2.5 32B Instruct (Studio, no surcharge, Apache 2.0)

Mature 30B-class baseline, heavily documented, lots of community recipes. The safer choice if you want 30B-class capability without betting on a new release.

Qwen3 32B (Studio, no surcharge, Apache 2.0)

Qwen 3 dense flagship at 32B — thinking-mode reasoning, heavily documented community recipes. Apache 2.0. Pick over Qwen 2.5 32B for the newer generation; pick 2.5 32B for more-documented recipes.

DeepSeek-R1-Distill-Qwen-32B (Studio, +$200 surcharge, MIT)

Reasoning-distilled 32B. Compare it on task-specific answer-quality evaluations; the surcharge reflects additional training scope, not preservation of hidden chain-of-thought traces.

Qwen3 30B-A3B MoE (Studio, +$300 surcharge, Apache 2.0)

Qwen 3 sparse MoE — 30B total, 3B active per token. Sparse activation reduces per-token compute but the full weights still need memory or offload; consumer-GPU deployment generally requires quantization and runtime-specific testing.

Qwen 3.5 35B-A3B MoE (Studio, +$300 surcharge, Apache 2.0)

Sparse MoE — 35B total, 3B active. Strong reasoning with thinking mode. Surcharge for MoE memory overhead; the full parameter count stays resident even under sparse activation during training.

Qwen 3.6 27B (Studio, included, Apache 2.0)

Latest Qwen dense 27B release, announced in April 2026. Best when you want the newer Qwen 3.6 agentic coding behavior without taking on MoE training overhead or a model surcharge.

Qwen 3.6 35B-A3B MoE (Studio, +$300 surcharge, Apache 2.0)

Latest Qwen MoE, April 2026 release with agentic coding upgrades. Pick this over 3.5 35B-A3B for newer agentic behavior; pick 3.5 if you want the more-documented generation.

Qwen 3.8 27B (Studio, included, Apache 2.0)

Official August 2026 Qwen dense release with native vision-language input and 262K native context. It is the current Qwen reference in the 27B class. Because the release is new, Logos confirms trainer, multimodal-data, and target-runtime compatibility during intake rather than implying every workflow is already production-proven.

Hermes 4.3 36B (Studio, +$300 surcharge, Apache 2.0)

Nous Research's 36B Hermes release uses an Apache-2.0 license according to its official model card. Confirm context, runtime, and benchmark behavior from that card rather than transferring undocumented “70B-class” guarantees.

Mixtral 8x7B Instruct (Studio, no surcharge, Apache 2.0)

The catalog entry omitted from the previous 44-item article. It has roughly 47B total parameters and 13B active per token; sparse compute does not remove the need to hold or offload the full weights.

49B — NVIDIA's instruction-tuned flagship

Llama-3.3-Nemotron-Super-49B (Studio, +$300 surcharge, NVIDIA Open Model License)

NVIDIA's instruction-tuned Llama with exceptional alignment and tool-calling reliability. The surcharge reflects the care required to preserve NVIDIA's alignment scaffolding during fine-tuning — disturb it carelessly and you lose what makes the model interesting.

Gemma 4 frontier — vision-language

Gemma 4 26B-A4B MoE (Studio, no surcharge, Apache 2.0)

Gemma 4's 26B-A4B model supports image input, but not audio, and has a 256K-token context window according to the official model card. Confirm serving support at intake.

Gemma 4 31B Dense (Studio, no surcharge, Apache 2.0)

The dense 31B variant supports image input, not audio, and has a 256K-token context window. Any Arena rank needs a named leaderboard, snapshot date, and evaluation settings, so the previous unsourced “#3” claim has been removed.

70B–141B — large dense and sparse models

Llama 3.3 70B Instruct (Atelier, no surcharge, Llama 3.3 license)

A widely used 70B-class instruction baseline listed at the Atelier base price. Production suitability depends on task evaluation, deployment controls, hardware, and compliance with the Llama 3.3 license.

Qwen 2.5 72B Instruct (Atelier, no surcharge, Qwen license)

An established 72B baseline to evaluate for multilingual and mathematics-heavy tasks. Compare it directly with Llama 3.3 70B on the intended dataset and runtime. Its license includes a condition for organizations above 100 million monthly active users.

Hermes 3 Llama 3.1 70B (Atelier, no surcharge, Llama 3.1 license)

Hermes 3 on a Llama 3.1 70B base, tuned by Nous Research for assistant behavior and function calling. Compare its existing prompt/tool behavior with a vanilla instruction model before deciding which needs less adaptation for the target task.

Mixtral 8x22B Instruct (Atelier, no surcharge, Apache 2.0)

An Apache-2.0 sparse model with roughly 141B total parameters and 39B active per token, listed at the Atelier base price. Validate coding and reasoning quality, memory placement, throughput, and runtime support on the target system.

Frontier MoE — 122B total-parameter class and up

The Atelier tier with surcharge. These models require capacity review: total weights, optimizer state, activations, quantization, offload, sequence length, and adapter method determine whether a single node is feasible.

Qwen 3.5 122B-A10B MoE (Atelier, +$799 surcharge, Apache 2.0)

Qwen 3.5 Medium (Feb 2026). The official model card identifies 122B total parameters and 10B active parameters; the 122B figure is not the active count. Hardware and throughput still depend on the full artifact, precision, context, and runtime.

MiniMax-M2 230B-A10B MoE (Atelier, +$1,499 surcharge, MIT)

Agent-native frontier MoE — 230B total, 10B active. Optimized for coding and agentic tool-calling; open-sourced Oct 2025. Compact-and-fast alternative to M1 when agentic behavior is the dominant target.

Qwen3 235B-A22B MoE (Atelier, +$1,499 surcharge, Apache 2.0)

Qwen3's frontier MoE from mid-2025 — 235B total, 22B active per token. Hardware and duration are quoted after a memory plan and a measured pilot; 8× H200 and 12 hours are not universal minima.

Llama-3.1-Nemotron-Ultra-253B (Atelier, +$1,799 surcharge, NVIDIA Open Model License)

NVIDIA's 253B reasoning model. Exact throughput, memory, and LoRA feasibility depend on precision, sequence length, parallelism, offload, and implementation; no fixed 8× H200 guarantee is made here.

Qwen 3.5 397B-A17B MoE (Atelier, +$2,499 surcharge, Apache 2.0)

Current Qwen flagship MoE — 397B total, 17B active. Native multimodal agent capabilities. Multi-day training potential for full LoRA convergence.

MiniMax-M1 456B-A45.9B MoE (Atelier, +$2,499 surcharge, Apache 2.0)

The first open-weight large-scale hybrid-attention reasoning MoE. 456B total, 45.9B active with lightning attention. Native 1M-token context (8× DeepSeek-R1). Multi-node H200 cluster with careful attention-kernel preservation during training.

GLM-5 744B MoE (Atelier, +$2,999 surcharge, MIT)

Z.ai's flagship MoE — 28.5T pre-training tokens, best-in-class open-source reasoning + coding + agentic performance. The base model that powers the GLM-5.1 series.

GLM-5.1 754B MoE (Atelier, +$3,499 surcharge, MIT)

Z.ai's official GLM-5.1 model card reports 58.4 on SWE-Bench Pro under its evaluation setup. Treat that as a vendor-reported snapshot, not a timeless “#1” or a general claim that it beats every closed model. Training scope is confirmed after capacity review.

Kimi K2.5 1.04T MoE (Atelier, +$4,499 surcharge, Modified MIT)

Trillion-parameter frontier MoE with vision and agent-oriented modes. The previous “coding parity with GPT-5 and Gemini” claim lacked a defined evaluation. Quantization availability and multi-node feasibility are confirmed during intake for the exact artifact and runtime.

How to pick

The practical decision tree most teams end up at:

"I want a widely deployed 70B-class baseline" → evaluate Llama 3.3 70B (Atelier). If you need function-calling behavior in the base, also evaluate Hermes 3 Llama 3.1 70B.
"I want 70B class but multilingual or math-heavy" → Qwen 2.5 72B (Atelier).
"I want strong reasoning at small size for inference cost reasons" → Phi-4 14B (Studio) or Qwen3 14B (Starter).
"I need multimodal input" → verify the exact modality: Gemma 4 E4B supports vision and audio, while the 26B and 31B variants support vision but not audio.
"I need long-context (128K+)" → Gemma 4 E2B and E4B provide 128K context, while Gemma 4 12B, 26B-A4B, and 31B provide 256K. Qwen 3.5, Hermes 4.3 36B, and MiniMax-M1 limits must likewise be confirmed from the exact model card and runtime before intake.
"I want agentic coding at the frontier, cost is secondary" → GLM-5.1 or Kimi K2.5 (Atelier + surcharge). MiniMax-M2 is the agentic-efficiency alternative.
"I'm prototyping, minimize spend" → Llama 3.1 8B or Qwen3 8B (both Starter).
"My workload is reasoning-heavy and I want to fine-tune the best reasoning distillate" → DeepSeek-R1-Distill-Qwen-32B (Studio +$200).
"I want a model with function-calling already trained in" → Hermes 3 Llama 3.1 8B (Starter) or Hermes 3 Llama 3.1 70B (Atelier).
"I want the longest possible context at the frontier" → MiniMax-M1 (1M tokens, Atelier + $2,499).

Licensing notes

Many catalog entries use Apache-2.0 or MIT licenses, which generally permit commercial use subject to their notices, conditions, and any other applicable data, trademark, export, or product obligations. Other entries use model-specific terms. Examples that require exact-license review include:

Llama 3.1 / 3.3 licenses — include attribution and an additional permission requirement for products above the license's 700M monthly-active-user threshold. Applies to vanilla Llama and Llama-based Hermes 3 tunes; read the current license.
NVIDIA Open Model License — applies to the Llama-Nemotron Nano, Super, and Ultra variants. Commercial use and derivative works are allowed under NVIDIA's published terms.
Gemma 4 — released under Apache-2.0 according to Google's Gemma 4 model card. Older Gemma generations use Google's Gemma terms instead; do not transfer those older terms onto Gemma 4 or assume Apache-2.0 applies retroactively.
Qwen license (Qwen 2.5 72B specifically) — restricts use by companies with over 100M monthly active users.
Modified MIT (Kimi K2.5) — broad MIT with minor usage restrictions; check the full text if you're operating at scale.
Hermes 4.3 36B — the official model card identifies Apache-2.0; verify the exact artifact and bundled notices at intake.

The Logos model selector surfaces the license per model. If you have licensing constraints, pre-filter by Apache 2.0 or MIT.

What Logos does with your chosen model

Once scope is accepted, Logos trains the selected artifact and delivers the agreed weights or adapter plus an evaluation report. Your rights remain subject to the base-model license, dataset rights, and engagement terms. Runtime compatibility and conversion requirements are documented per delivery.

Open the Logos model selector to see pricing per base model →

For the decision between Logos and the big-name alternatives, see Logos vs OpenAI vs Together vs Replicate. For the product overview, What is Logos?.