If you want one name for free coding in 2026, take DeepSeek V4 Pro. It works without a credit card on chat.deepseek.com, its context window reaches 1 million tokens, its weights are released under the MIT license, and it resolves 80.6% of SWE-bench Verified tasks according to its official model card. No other free model in this comparison checks those four boxes at the same level.

The ranking changes as soon as you leave the general case. Kimi K2.6 is the strongest option for long agentic sessions, with 80.2% on SWE-bench Verified and 66.7% on Terminal-Bench 2.0. GLM-4.7 gives you a free API through GLM-4.7-Flash and fits well inside Claude Code, Cline or Roo Code. Qwen3-Coder is the best pick if you want to self-host an Apache-2.0 model. Gemini 3.8 Flash is the easiest free model to call from the Gemini API. GitHub Copilot Free and OpenRouter complete the picture, one inside the editor and the other as a shared gateway.

We checked official pages, model cards and pricing tables in November 2026. Prices move fast, and model versions move faster. Every figure below comes from a page we opened; when a vendor does not publish a benchmark, we say so instead of filling the gap.

The comparison at a glance#

CriterionDeepSeek V4 Pro Notre choixKimi K2.6GLM-4.7Qwen3-Coder 480BGemini 3.8 Flash
Free accessChat on chat.deepseek.comKimi chat, $0 Free planz.ai chat and GLM-4.7-Flash APIChat on chat.qwen.aiGemini API free tier
Weight licenseMITModified MITMITApache-2.0Proprietary
Parameters1.6T total, 49B active1T total, 32B active355B total, 32B active480B total, 35B activeNot disclosed
Context1M tokens256K tokens200K tokens256K native, 1M with YaRNNot listed on the pricing page
SWE-bench Verified80.6%80.2%73.8%Not listed on the model cardNot listed
Terminal-Bench 2.067.9%66.7%41.0%Not listedNot listed
API price per million tokens$0.66 input / $1.98 output off-peak¥6.50 input / ¥27 output$0.60 input / $2.20 outputPaid through Alibaba Model Studio$0 on the free tier, then $0.75 / $3.75
Main limitPaid API, no guarantee on free chatK2 retires at the end of December 2026The Flash variant is much smallerSelf-hosting is very heavyFree-tier data is used to improve products
Best forThe best overall free modelAgents that run for hoursA free API and agent workflowsApache-2.0 self-hostingPrototypes inside Google's ecosystem
Prices and specifications checked in November 2026 on official pages. Benchmarks are vendor-reported and do not all measure the same thing.

How we decided#

Five criteria drive the score: price and access without a card, code quality on public benchmarks, behaviour on long agentic sessions, how open the weights are, and privacy terms. These are editorial scores, not performance measurements. A model that wins a benchmark does not automatically win your project: a 40,000-line repository, a TypeScript monorepo and a one-off migration script have different needs.

ToolPrice and free accesspoids 30 %Code qualitypoids 25 %Context and agentspoids 20 %Weight opennesspoids 15 %Privacy and termspoids 10 %Score
DeepSeek V4 Pro
9
9
9
10
7
9,0
Kimi K2.6
8
9
10
8
6
8,5
GLM-4.7
9
7
8
9
7
8,1
Qwen3-Coder
8
7
8
10
8
8,1
Gemini 3.8 Flash
8
8
8
1
4
6,6
Final score = weighted average of the criteria, out of 10.

DeepSeek wins on balance, Kimi on agentic depth, GLM on API price, Qwen on deployment freedom and Gemini on ease of access. That is why the final verdict is a set of profiles rather than a single winner for every reader.

DeepSeek V4 Pro: the best power-to-price ratio#

DeepSeek V4 Pro is a Mixture-of-Experts model with 1.6 trillion parameters, 49 billion of them active per token. Its context window reaches 1 million tokens and its maximum output is 384,000 tokens, according to the API documentation. The model card reports 80.6% on SWE-bench Verified, 67.9% on Terminal-Bench 2.0 and 76.2% on SWE-bench Multilingual. Those are DeepSeek's own numbers, but they put the model in the same conversation as much more expensive closed models.

Free access runs through the official chat. You sign in, pick a reasoning mode and start working. The chat does not ask for a card and accepts code files within its context window. For a solo developer, it is the fastest way to test the model on a real repository before paying anything.

The API is paid, but it is one of the cheapest around. Off-peak, DeepSeek charges $0.66 per million input tokens (cache miss) and $1.98 per million output tokens. At peak, those prices rise to $1.32 and $3.96. Input tokens served from cache cost $0.022 off-peak and $0.044 at peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays. A migration script that runs at night or over the weekend therefore costs half as much.

The weak spot is local deployment. With 1.6 trillion parameters, the model has no business on an RTX 4090. Even quantized, it needs several accelerators and a lot of memory. Self-hosting is for teams that already own the infrastructure. For everyone else, the free chat or the paid API are the two realistic entry points.

DeepSeek V4 Pro at a glance

Strengths

  • 1M context and 384K maximum output
  • MIT weights, so commercial use is allowed
  • Very cheap API outside peak hours
  • 80.6% on SWE-bench Verified

Limits

  • Free chat has no availability guarantee
  • Self-hosting is out of reach for a desktop machine
  • Benchmarks are vendor-reported

Kimi K2.6: built for long agentic sessions#

Kimi K2.6 is a multimodal model with 1 trillion parameters, 32 billion of them active. Its context window is 256,000 tokens and its weights are released under a modified MIT license. Moonshot reports 80.2% on SWE-bench Verified, 66.7% on Terminal-Bench 2.0 and 58.6% on SWE-bench Pro. The model card says coding scores are averaged over ten independent runs.

Its argument is not the raw score but the duration. In the official announcement, Moonshot describes a session of more than 12 hours and 4,000 tool calls to optimize local inference in Zig, and another 13-hour run on an eight-year-old financial matching engine, with more than 1,000 tool calls and 4,000 lines changed. These are vendor-selected examples, not reproducible tests, but they describe the use case well: an agent that must keep the thread for hours and recover from its own mistakes.

Kimi is free to use in the chat, with a $0 Free plan. The paid Plus plan starts at $15 a month billed yearly, with more agent credits and Kimi Code access. The API is billed at ¥6.50 per million input tokens and ¥27 per million output tokens for K2.6, according to the official price list.

There is a calendar limit you need to know. Moonshot closed new sign-ups for K2 models on October 31, 2026, and plans to retire the entire K2 series at the end of December 2026. If you build a workflow on K2.6 today, plan the migration to K3 or to self-hosted weights. The free chat is useful for testing, but it is not a foundation for a product that must still run in 2027.

Kimi K2.6 at a glance

Strengths

  • Excellent on long agentic tasks
  • 256K context and vision input
  • Modified MIT weights
  • $0 Free plan

Limits

  • The K2 series disappears at the end of December 2026
  • API priced in Chinese yuan
  • Long-run examples come from the vendor

GLM-4.7: the free API and the agent integration#

GLM-4.7 is Z.ai's model for code and agents. It has 355 billion parameters, 32 billion active, a 200,000-token context window and MIT-licensed weights. On the vendor's benchmarks it reaches 73.8% on SWE-bench Verified, 66.7% on SWE-bench Multilingual and 41.0% on Terminal-Bench 2.0. It trails DeepSeek and Kimi on the first two scores, but it remains competitive for editing tasks and agent workflows.

Its most concrete advantage is the free API. Z.ai offers GLM-4.7-Flash, a 30B-A3B model, with input and output priced at $0 on the pricing page. This is not the full GLM-4.7: Flash is smaller, faster and less capable on very long tasks. But for test generation, local refactoring, code translation or a background agent, it is a free option that does not require a credit card.

The full GLM-4.7 costs $0.60 per million input tokens and $2.20 per million output tokens. It is available in the chat at chat.z.ai, and Z.ai documents integrations with Claude Code, Cline, Roo Code and Kilo Code. Its "Preserved Thinking" mode keeps reasoning blocks across turns, which saves the agent from rebuilding its analysis at every step. That matters on a large repository where the model must remember why it rejected an approach.

The limit is transparency. The name "GLM-4.7" covers several versions, and Flash does not perform like the full model. Check the exact model name in your configuration, especially if you use a gateway that mixes versions. And like every free tier, the $0 price can change: nothing guarantees that GLM-4.7-Flash will still be free in six months.

GLM-4.7 at a glance

Strengths

  • Free GLM-4.7-Flash API
  • MIT weights and a 200K context
  • Good integration with Claude Code, Cline and Roo Code
  • Preserved Thinking for long tasks

Limits

  • Flash is much smaller than the full model
  • Version names are easy to confuse
  • Benchmarks are vendor-reported

Qwen3-Coder: the Apache-2.0 choice for self-hosting#

Qwen3-Coder 480B-A35B is a Mixture-of-Experts model with 480 billion parameters, 35 billion of them active. It supports 256,000 tokens of context natively and up to 1 million with YaRN, according to the official model card. Its Apache-2.0 license is the most permissive in this group: you can deploy it, modify it and ship it inside a product without a non-compete clause or a redistribution restriction.

Its performance is harder to compare. Qwen publishes 38.7% on SWE-bench Pro, a different test from the SWE-bench Verified scores used by DeepSeek, Kimi and GLM. That number does not belong in the same column. The model is still strong on generation and completion tasks, and it is well tooled: a dedicated function-call format, support for Qwen Code, Cline and the usual inference servers.

Free access runs through Qwen Chat. Qwen Code, however, is no longer a free CLI with a simple account: the official documentation now asks for a Qwen Plus or Pro subscription, or an API key. The Qwen Code repository is open source under Apache-2.0 and shows 29,000 stars, but model access through the CLI is billed. That distinction matters: the agent code is free, the model that runs it is not.

Self-hosting is the real argument. With 480 billion parameters in BF16, you need roughly 960 GB of weights, which means several H100s or an equivalent cluster. This is not a model for a home server. But for a company that wants to keep its code on its own infrastructure, Apache-2.0 is the most reassuring license. If you deploy on a VPS, our hosting comparison for developers covers the plans that include SSH and Git, and our Coolify guide shows how to organize the deployment.

Qwen3-Coder at a glance

Strengths

  • Apache-2.0, the most permissive license here
  • 256K native and 1M with YaRN
  • Good function-calling tooling
  • Free chat on chat.qwen.ai

Limits

  • 480B parameters: self-hosting is very heavy
  • Qwen Code requires a subscription or API key
  • No SWE-bench Verified score on the model card

Gemini 3.8 Flash: the easiest free API#

Gemini 3.8 Flash is the free model in the Gemini API tier. Google states on its pricing page that input and output are free on the free tier, with an explicit trade-off: content sent to the free tier can be used to improve Google's products. The paid model costs $0.75 per million input tokens and $3.75 per million output tokens until December 31, 2026, then $1.50 and $7.50 from January 1, 2027.

It is the simplest choice if you already live in Google's ecosystem. One API key, one call compatible with the usual SDKs, and you have a coding model for a prototype, a script or a small app. The free tier does not require a credit card, but the quotas are not published in detail on the pricing page: they show up in AI Studio and depend on the project and the model. Do not assume a fixed number of requests per day without checking your own account.

Watch out for a common confusion. The free Gemini CLI with a personal Google account was replaced by Antigravity CLI on June 18, 2026, according to the Gemini CLI documentation. The often-quoted 1,000 requests per day no longer describes the current free access. The Gemini API still has a free tier, but it is not the same door as the old CLI.

The main limit is privacy. A proprietary model, a free tier that reuses data and quotas that can change: for client code under a confidentiality agreement, move to the paid tier or choose a model you host yourself. Gemini 3.8 Flash is excellent for learning, prototyping and automating non-sensitive tasks. It is much less suitable for a private repository that must not leave your infrastructure.

Gemini 3.8 Flash at a glance

Strengths

  • Free tier without a credit card
  • Simple, well-documented API
  • Good for prototypes and multimodal tasks
  • Competitive paid price until the end of 2026

Limits

  • Free-tier data is used to improve products
  • Free quotas are not published in detail
  • Proprietary model, no self-hosting

GitHub Copilot Free: 2,000 completions a month in your editor#

GitHub Copilot Free is not a model; it is an access path. The free plan includes 2,000 completions per month, a monthly allowance of GitHub AI credits and automatic model selection, according to the GitHub documentation. It also gives you Copilot Chat in the IDE, on GitHub.com and in GitHub Mobile, plus the Copilot CLI and app.

It is the simplest choice for a beginner or for a developer who wants completion inside VS Code without configuring an API key. Install the extension, sign in with your GitHub account and start coding. The free plan does not let you choose the model: GitHub decides for you. That is a black box, but it is enough for scripts, exercises and small projects.

The limit arrives quickly on a real project. 2,000 completions per month is about 65 per working day. A developer who codes eight hours a day with autocomplete enabled can burn through the quota in a few days. Copilot Pro at $10 a month gives 1,500 AI credits and access to a selection of models; Pro+ at $39 goes up to 7,000 credits. If you need to choose Claude, GPT or Gemini inside the editor, look elsewhere.

GitHub Copilot Free at a glance

Strengths

  • 2,000 completions per month at no cost
  • Direct integration with VS Code, JetBrains and GitHub
  • No API key to configure

Limits

  • Automatic model selection only
  • Quota runs out fast on an intensive project
  • No access to premium models

OpenRouter: the gateway to free models#

OpenRouter is not a model either, but a gateway that gives you access to several models with one API key. Free variants, marked with the :free suffix, are limited to 20 requests per minute and 50 requests per day if you have purchased less than $10 in credits. If you have purchased at least $10 once, the cap rises to 1,000 requests per day, according to the OpenRouter documentation. The $10 is not a subscription: it stays on your account.

This is the ideal tool for comparing several models without creating five accounts. You can send the same prompt to DeepSeek, Qwen, GLM or Kimi and see which one produces the cleanest diff. The gateway also handles providers, fallbacks and capacity errors. For a prototype, it is faster than integrating each API separately.

The limit is production. Free models share provider capacity and can be slowed down or refused at peak times. The 50-requests-per-day cap is too low for an agent running continuously. And free models change: a model can disappear from the catalogue overnight. Use OpenRouter for testing, not to guarantee a service.

OpenRouter at a glance

Strengths

  • One key for several models
  • Free models without a subscription
  • 1,000 requests per day after a one-time $10 purchase

Limits

  • 20 requests per minute on free models
  • Capacity shared with other users
  • Catalogue and free models change often

What the prices don't tell you#

The first trap is quotas. A "free" model is not an unlimited model. DeepSeek and Kimi limit free chat capacity at peak times. Gemini applies per-project quotas that are not published on the pricing page. OpenRouter caps free models at 20 requests per minute. GitHub Copilot Free stops after 2,000 completions. Before building a workflow, read the limits page, not just the price page.

The second trap is context. A 1-million-token window does not mean the model works well on 1 million tokens. Long-context benchmarks often measure a precise retrieval inside a text, not the ability to understand an entire repository. A model that answers correctly at 200,000 tokens can lose the thread at 800,000. Test on your own code before trusting the marketing number.

The third trap is licensing. "Open weights" does not mean "open source". DeepSeek and Qwen release weights under MIT and Apache-2.0, which allows commercial use. Kimi uses a modified MIT with additional conditions. GLM-4.7 is MIT, but Flash and the full model do not have the same capabilities. Gemini and GitHub Copilot are proprietary. If you resell a product, check the weights license, not only the agent's code license.

The fourth trap is data. Gemini's free tiers use content to improve products. Free chats from DeepSeek, Kimi and Qwen have their own terms, which can change by country and account. For client code, a self-hosted model or a paid API with a data processing agreement is safer. Free sometimes costs you privacy.

The fifth trap is end of life. Kimi K2.6 disappears at the end of December 2026. Free models change names, prices and limits several times a year. An article dated November 2026 can be wrong by March 2027. Keep an abstraction in your code: an OpenAI-compatible client that can point to DeepSeek, Qwen, GLM or Gemini without rewriting your application. That is the only way to avoid depending on one model.

Do you need a VPS to code with a free model?#

Not necessarily. The models in this article work well through their chat or API, with no server to manage. A small VPS becomes useful if you want to run an agent continuously, host a CI runner or isolate a model from the rest of your machine. A small server is enough to start, and our hosting comparison covers the plans that include SSH and Git.

For a local model, the math is different. Qwen3-Coder 480B or DeepSeek V4 Pro need several accelerators, not a $10-a-month VPS. A small model such as GLM-4.7-Flash or a mid-sized Qwen can run on a machine with a recent GPU. If you document team conventions, a self-hosted wiki such as Docmost pairs well with an AGENTS.md file: the rules stay readable by humans and agents. Our Docker Compose tutorial for Docmost walks through the install in fifteen minutes.

Our pick, by profile#

Our pick

DeepSeek V4 Pro

8.9/10

DeepSeek V4 Pro wins the overall ranking: free in the chat, a 1M context, 80.6% on SWE-bench Verified and a very cheap API. Kimi K2.6 remains our pick for long agentic sessions, GLM-4.7-Flash for a free API, Qwen3-Coder for Apache-2.0 self-hosting and Gemini 3.8 Flash for prototypes inside Google's ecosystem.

Who it's for
Solo developers who want the strongest free model without a credit card, with a 1M context and MIT weights.
Who should skip it
Anyone who needs an agent to run unattended for more than twelve hours: Kimi K2.6 is stronger, provided you plan for its end of life.

Frequently asked questions#

Frequently asked questions

What is the best free coding model in 2026?

DeepSeek V4 Pro for the general case: free on chat.deepseek.com, a 1M context, 80.6% on SWE-bench Verified and MIT weights. Kimi K2.6 is better on long agentic sessions, GLM-4.7-Flash on a free API and Qwen3-Coder on self-hosting.

Can companies use these models for free?

It depends on the license and the tier. DeepSeek, GLM and Qwen release weights under MIT or Apache-2.0, so self-hosting is possible. Kimi uses a modified MIT. Gemini and GitHub Copilot are proprietary. Gemini's free tier uses data to improve products: for client code, move to the paid tier or self-host.

Is Kimi K2.6 really going away?

Yes, the K2 series is at the end of its life. Moonshot closed new sign-ups on October 31, 2026, and plans to retire all K2 models at the end of December 2026. The weights remain downloadable and self-hostable, but chat and API access to K2.6 will not last.

What is the difference between open weights and open source?

An open-weight model publishes its parameters, but not necessarily the training code, the data or the full license. DeepSeek and Qwen are close to open source with MIT and Apache-2.0. Kimi adds conditions. Gemini and Copilot do not publish weights. Read the license before reselling a product.

Does Gemini 3.8 Flash require a credit card?

No, the Gemini API free tier does not require a credit card. Quotas are limited, and content can be used to improve Google's products. Exact limits appear in AI Studio and depend on the project.

Which model fits a small GPU or a home server?

None of the large models in this article fits on an RTX 4090. DeepSeek V4 Pro, Kimi K2.6, GLM-4.7 and Qwen3-Coder 480B need several accelerators. For a small machine, look at GLM-4.7-Flash, a mid-sized Qwen or a quantized 7B to 30B model.

Sources#

Official pages checked on November 6, 2026.