Best Open Source LLMs in 2026: We Reviewed 7 Models

Best Open Source LLMs in 2026: We Reviewed 7 Models

Announcing our Series D and $1B ARR

Blog

Best Open Source Llms

Best Open Source LLMs in 2026

PUBLISHED 7/8/2026

Table of Contents

The best open source LLMs at a glance

What makes a great open source LLM?

How we evaluated these models

How these models compare on benchmarks

Reasoning and knowledge

Software engineering performance

Key takeaways

GLM 5.2

What is GLM 5.2?

Who should use GLM 5.2?

Standout features

Pros and cons

FAQ

Kimi K3

What is Kimi K3?

Who should use Kimi K3?

Standout features

Pros and cons

FAQ

Kimi K2.7 Code

What is Kimi K2.7 Code?

Who should use Kimi K2.7 Code?

Standout features

Pros and cons

FAQ

DeepSeek-V4-Pro

What is DeepSeek-V4-Pro?

Who should use DeepSeek-V4-Pro?

Standout features

Pros and cons

FAQ

DeepSeek-V4-Flash

What is DeepSeek-V4-Flash?

Who should use DeepSeek-V4-Flash?

Standout features

Pros and cons

FAQ

MiniMax M3

What is MiniMax M3?

Who should use MiniMax M3?

Standout features

Pros and cons

FAQ

Qwen3.7 Plus

What is Qwen3.7 Plus?

Who should use Qwen3.7 Plus?

Standout features

Pros and cons

FAQ

gpt-oss-120b

What is gpt-oss-120b?

Who should use gpt-oss-120b?

Standout features

Pros and cons

FAQ

Gemma 4 31B IT

What is Gemma 4 31B IT?

Who should use Gemma 4 31B IT?

Standout features

Pros and cons

FAQ

Why run these models on Fireworks

Performance at scale

Fireworks deployment options

With new open source LLMs launching constantly, figuring out which model actually fits your use case has become its own endeavor.

The open-source model landscape is moving fast in 2026. GLM 5.2, Kimi K3 (described by Kimi as open source, though its full weights pending, scheduled for released by July 27, 2026), Kimi K2.7 Code, MiniMax M3, and DeepSeek-V4-Pro all launched between April and July, and several now sit within a few benchmark points of frontier models at a fraction of the serving cost.

The right model depends on four constraints: benchmark quality, context window, modality support, and license terms. A model that leads on composite quality but lacks image input loses to a weaker model when screenshots or visual documents are part of the product. A model with a trillion-parameter MoE architecture and a 1,040k-token context window is the wrong default for a high-throughput extraction route where a smaller, faster model finishes the job at lower cost.

TL;DR

The table below spans models for general reasoning, coding, multimodal, long-context, and high-throughput workloads.

Kimi K3 has the highest Artificial Analysis Intelligence and Coding Index scores among the nine models evaluated here. Its weights and license are still pending.

GLM 5.2 leads both indexes among open models available on Fireworks. Other models remain competitive when modality, context window, licensing, or serving cost sets the constraint.

Use the benchmark tables to narrow the field, then run task-level evals before choosing a production route.

The best open source LLMs at a glance

ModelRelease dateParamsContext windowBest forOn Fireworks

GLM 5.2June 2026743B total1,040k tokensStrongest benchmark profile among currently open models in this set. First eval for broad reasoning, coding, and long-context agents.Try in playground

Kimi K3July 20262.8T total, 16 of 896 experts active1M tokensBenchmark leader awaiting full weights and a Fireworks listing. License details remain unpublished.Not yet listed

Kimi K2.7 CodeJune 20261.02T total262k tokensCoding agents, repository work, patch planning, and multimodal developer tools. Thinking mode is mandatory.Try in playground

DeepSeek-V4-ProApril 20261.6T total1,040k tokensLong-context reasoning and coding from a separate open-source family. Second eval when GLM 5.2 misses on your repo.Try in playground

DeepSeek-V4-FlashApril 2026284B total1,040k tokensSame 1,040k context class as Pro at higher throughput and lower cost. Default DeepSeek route for high-volume workloads.Try in playground

MiniMax M3June 202622, 2025428B total, about 23B activated512k tokensNative image and video input, second-highest GPQA score in the table. First eval when multimodality sets the constraint.Try in playground

Qwen3.7 PlusJune 2026N/A262k tokens Closed-weight Qwen-family model with image input and training support through Fireworks. Review license before production.Try in playground

gpt-oss-120bAugust 2025116B total131k tokensApache-2.0 licensing and fastest median output in this set at 271.4 tokens/second. For bounded, high-throughput tasks.Try in playground

Gemma 4 31B ITApril 202632.2B dense262k tokensSmaller dense multimodal model with Apache-2.0 licensing and an On-Demand deployment path. First eval when adaptation and deployment control matter more than frontier benchmark rank.View On-Demand model

What makes a great open source LLM?

Architecture helps narrow the evaluation set. A trillion-parameter MoE model can route each token through a smaller active subset, but parameter counts do not establish serving cost or latency. Measure both on the route and workload you plan to use.

Context window size counts only when the model holds coherence at length. A million-token-class context window is useful for codebases, large collections of legal documents, retrieval-heavy agents, and document workflows. A million-token context window only helps if the model can stay coherent for the entire prompt. If it loses track of earlier instructions or evidence halfway through, you end up paying for extra context that does not improve the output.

Licensing and access to the model weights determine whether you can use the model in production. Apache-2.0 and MIT are easier for most teams to approve.

Not all “open source” model licenses are equally simple to use in production. Licenses like Modified MIT and the MiniMax Community license can still be acceptable, but they usually come with extra conditions, so they should be reviewed before you commit to a model. Qwen3.7 Plus is closed weights, so legal and training terms should go through legal review before the model is evaluated.

Model selection is not only about benchmarks. Your deployment path (Serverless or dedicated) determines which models fit your latency, cost, and capacity requirements.

Serverless works for fast trials and bursty traffic. On-Demand fits dedicated capacity, stricter latency targets, and predictable production volume. If you keep repeating the same constraints in prompt after prompt, that is the signal to add post-training to the rollout plan.

How we evaluated these models

When we compared these open source LLMs, we focused on the factors that most directly determine whether a model will work for a real workload: benchmark performance, context window length, license or weight access, multimodal support, and the serving path you plan to run in production.

How these models compare on benchmarks

The benchmark mix covers broad intelligence, hard science reasoning, coding ability, scientific coding, and output speed. Artificial Analysis Intelligence Index and Coding Index identify first-pass candidates. Use GPQA, HLE, and SciCode to find where a model breaks under harder prompts.

Reasoning and knowledge

ModelAAIIGPQAHLE

Kimi K357.193.5%44.3%

GLM 5.251.189.5%40.1%

MiniMax M344.492.9%37.1%

DeepSeek-V4-Pro44.388.8%35.9%

Kimi K2.7 Code41.989.6%32.8%

DeepSeek-V4-Flash40.389.4%32.1%

Qwen3.7 Plus39.090.0%33.4%

Gemma 4 31B IT29.485.7%22.7%

gpt-oss-120b23.878.2%18.5%

Claude Fable 5 59.992.6%53.3%

Claude Opus 4.8 55.792.0%45.7%

What reasoning and knowledge benchmarks measure

•Artificial Analysis Intelligence Index: A composite score across hard model evaluations, including GPQA, HLE, SciCode, AA-LCR, and other tasks.

•GPQA: Graduate-level science reasoning.

•HLE: Humanity's Last Exam, a hard knowledge and reasoning benchmark where most models still have low absolute scores.

Software engineering performance

ModelAACISciCodeMedian output speed

Kimi K376.258.7%58.5 tok/sec

GLM 5.268.850.5%165.3 tok/sec

Kimi K2.7 Code60.847.5%43.2 tok/sec

DeepSeek-V4-Pro59.450.0%60.1 tok/sec

MiniMax M358.645.4%91.7 tok/sec

DeepSeek-V4-Flash56.244.9%97.9 tok/sec

Qwen3.7 Plus55.945.5%53.0 tok/sec

Gemma 4 31B IT43.443.4%35.9 tok/sec

gpt-oss-120b30.438.9%271.4 tok/sec

Claude Fable 5 76.560.2%56.5 tok/sec

Claude Opus 4.8 74.353.5%54.1 tok/sec

What software engineering benchmarks measure

•Artificial Analysis Coding Index: A composite coding score for software engineering tasks.

•SciCode: Scientific coding and reasoning tasks.

•Median output speed: The median generated-token rate. Actual serving speed varies by model, route, prompt shape, and workload.

Scores current as of July 18, 2026. Actual performance varies by prompt, quantization, and inference settings.

Key takeaways

•Kimi K3 has the highest Intelligence and Coding Index scores among the nine models in the table. GLM 5.2 is the highest-scoring open model available on Fireworks today.

•GLM 5.2 is the first open-model quality eval across Intelligence Index and Coding Index. Fable 5 is the stronger composite reference, while Opus 4.8 provides continuity with earlier frontier comparisons.

•DeepSeek-V4-Pro and DeepSeek-V4-Flash give you two DeepSeek 1,040k-context routes: Pro for quality and Flash for generated-token rate.

•MiniMax M3 is the multimodal pick: second-highest GPQA score here, 512k context, and native image and video support.

•Qwen3.7 Plus is the closed-weights Qwen-family candidate when Fireworks access, image input, 262k context, and supported training workflows matter.

•gpt-oss-120b is the speed and Apache-2.0 route, but it trails the 2026 open frontier on composite quality.

GLM 5.2

What is GLM 5.2?

•Release date: June 2026

•Parameters: 743B total

•Context window: 1,040k tokens on Fireworks

•License: MIT license

•Corporate sponsor: Z.ai

•Model repo: zai-org/GLM-5

🚀 Try GLM 5.2 on Fireworks →

GLM-5.2 is Z.ai's flagship coding model, a 743B-parameter MoE with a 1,040k-token context window on Fireworks. Its IndexShare architecture reuses indexers across sparse attention layers and reduces per-token FLOPs by 2.9x at full context.

The GLM-5 line is built for long coding tasks and context-heavy tool use. GLM 5.2 adds a 1,040k context window, flexible reasoning effort, and a first eval fit for codebase context, multi-document retrieval, implementation planning, and tool-calling agents.

GLM 5.2 posts the highest composite scores of any model currently available on Fireworks in this table: 51.1 on the Intelligence Index and 68.8 on the Coding Index.

Who should use GLM 5.2?

Use GLM 5.2 as the default first eval for broad reasoning, coding, retrieval-heavy agents, and document workflows. It gives teams one baseline before they split work across specialized models.

GLM 5.2 is strongest when the workload needs both long context and function calling. If the request includes images, start with Kimi K2.7 Code, MiniMax M3, or Qwen3.7 Plus instead.

Standout features

•Highest Intelligence Index score among currently open, Fireworks-available models here: 51.1

•Highest Coding Index score among currently open, Fireworks-available models here: 68.8

•1,040k context window on Fireworks

•Function calling support

•Serverless and On-Demand availability on Fireworks

•165.3 tokens/second median output

Pros and cons

ProsCons

Strongest benchmark row among currently open, Fireworks-available models hereLarger model than most simple production tasks need

1,040k context for long documents and codebasesText-only; no image input

Function calling supportRequires task-level latency and cost testing against smaller routes

Serverless access makes evaluation fastText-only workloads should still be tested against multimodal alternatives when image input is likely

FAQ

Q: How much context should you test with GLM 5.2?

Start with the longest prompt your product can justify. Use a

[truncated for AI cost control]