benchr
Let's be exact about "free," because the word does a lot of work here. We're not talking about a free trial or a generous tier of a paid product. These are open-weight models. The weights are published under open licenses on Hugging Face, so you can download them and run them on your own hardware for nothing but compute. And each maker also runs a free hosted chat, so you can just use the model directly. That's a different thing from a paid IDE assistant, which wraps a model in tooling and bills you monthly. For that comparison, see the coding assistants shootout.
Three families lead the open-weight coding pack right now: DeepSeek out of China, Alibaba's Qwen, and Moonshot AI's Kimi. Here's how they stack up, by the scores the makers themselves publish.
| Model | Params (total / active) | License | Context | SWE-bench Verified |
|---|---|---|---|---|
| DeepSeek-V4-Pro | 1.6T / 49B | MIT | 1M | 80.6 |
| Kimi K2.6 | 1T / 32B | Modified MIT | 256K | 80.2 |
| DeepSeek-V4-Flash (weights only; hosted API retired Sep 10, 2026) | 284B / 13B | MIT | 1M | 79.0 |
| Qwen3.6-27B | 27B dense | Apache 2.0 | 262K | 77.2 |
| Qwen3.6-35B-A3B | 35B / 3B | Apache 2.0 | 262K | 73.4% |
| Qwen3-Coder-Next | 80B / 3B | Apache 2.0 | 256K | ~71% |
| Kimi K3 | 2.8T / 104B | Kimi K3 License | 1M | not published |
DeepSeek-V4-Pro: the score leader
By the strongest officially published number, DeepSeek-V4-Pro now tops this table. DeepSeek's model card reports 80.6 on SWE-bench Verified, the benchmark that scores real GitHub issue resolution, and 87.9 on Terminal-Bench 2.1. It's a 1.6-trillion-parameter Mixture-of-Experts model with 49 billion active per token, released in April 2026 under the MIT license with a 1-million-token context window. When this page first ran in May it printed no V4 score; benchr's record now carries DeepSeek's published figure, and the ranking here follows it.
The hosted side changed on September 10, 2026. DeepSeek retired the V4-Flash API model and replaced it with DeepSeek-V4.1-Flash; the old deepseek-v4-flash ID temporarily routes to the new model. V4-Flash's MIT weights stay downloadable, so its 79.0 on SWE-bench Verified still describes a model you can self-host. DeepSeek says V4.1-Flash beats V4-Pro on its benchmarks, but benchr has not recorded a figure or a license for it, so it stays out of the table. V4-Pro remains on the API: DeepSeek briefly announced it would route V4-Pro to V4.1-Flash from September 14, then reversed course and says V4-Pro stays available with the same billing.
DeepSeek's free chat at chat.deepseek.com is still the zero-setup way to try the family before you download anything; benchr has not rechecked which model it serves after the September 10 change. Where DeepSeek sits against the rest of the open field is covered in the open-weight tier right now.
Kimi K2.6, and why K3 isn't ranked
Moonshot AI's Kimi K2.6 sits 0.4 points behind, at 80.2 on SWE-bench Verified in Moonshot's model card, plus 58.6 on the tougher SWE-bench Pro and 76.7 on the multilingual variant. On SWE-bench Pro it is ahead of DeepSeek-V4-Pro's 55.4 (DeepSeek-reported), so which of the two leads depends on the benchmark you weigh.
Under the hood it's a 1-trillion-parameter Mixture-of-Experts model with 32 billion active per token, released in late April 2026 under a Modified MIT license. Moonshot leans hard into long-horizon agent work: K2.6 can coordinate up to 300 sub-agents across as many as 4,000 steps in a single run, with demonstrated 12-hour autonomous sessions. The catch is size. A trillion-parameter model is not something you spin up on a laptop, so for most people "free" here means the hosted chat or rented GPUs, not local. The full Kimi K2.6 review goes deeper on the Agent Swarm and what self-hosting costs.
Moonshot's newer model, Kimi K3, arrived on July 16, 2026, with a 1M-token context and API pricing of $3 input / $15 output per 1M tokens for Kimi K3. Moonshot publishes 88.3 on Terminal-Bench 2.1 for it, just above DeepSeek-V4-Pro's 87.9, plus 67.5 on DeepSWE and 93.5 on GPQA Diamond, but no SWE-bench Verified figure. That leaves no like-for-like number to rank it against K2.6 or DeepSeek on this table. Its weights ship under the Kimi K3 License rather than K2.6's Modified MIT, so read those terms before you build on it.
Qwen: the one you can run yourself
Qwen's pitch is efficiency, and it's the reason it might be the right answer even though it doesn't top the table. The Qwen3.6 generation from April 2026 gives you two options: Qwen3.6-27B, a dense model, scores 77.2 on SWE-bench Verified, and Qwen3.6-35B-A3B scores 73.4% while activating just 3 billion parameters per token out of 35 billion total. Alibaba also ships a dedicated coder, Qwen3-Coder-Next, an 80B model with the same 3B active footprint that lands a bit over 70% depending on the scaffold. All three are Apache 2.0, the most permissive license here, with no strings on commercial use. A newer Qwen3.8-27B checkpoint exists on Hugging Face and is not yet in benchr's record, so it isn't ranked here.
That small footprint is the headline. A 27B dense model or a 3B-active MoE runs on hardware that a trillion-parameter model can't touch, which puts a strong coder within reach of a single good GPU. If "free" has to mean "free on my machine," Qwen is the family to start with, and the small language models piece explains why the active-parameter trick works.
So which one
Go with DeepSeek-V4-Pro if you want the highest provider-published SWE-bench Verified score with open weights (80.6, MIT) and you can host it or use DeepSeek's API or chat. Kimi K2.6 is close enough at 80.2, and ahead on SWE-bench Pro, that its Agent Swarm or your own repository tests can decide between the two. Treat Kimi K3 as a candidate to test rather than a ranked pick until Moonshot publishes a SWE-bench Verified number. Go with Qwen if you need to run the model yourself on reasonable hardware: Qwen3.6-27B for the higher score, Qwen3.6-35B-A3B when you want only 3B parameters active per token. If you self-host DeepSeek-V4-Flash (79.0), remember its hosted API is gone; for hosted Flash-tier use, DeepSeek now serves V4.1-Flash.
One reality check before you commit: the two leaders are trillion-parameter models, so "free to download" and "free to run" are not the same sentence. If self-hosting is the goal, read running models on your own machine first, then size your ambitions to your GPUs.