Blog
flint vs qwen3.8flint coding modelqwen3.8 max

Flint vs Qwen3.8: Which Coding Model Should You Use?

Flint vs Qwen3.8 for coding: what each model is, which one you can run locally, which numbers are vendor claims, and how to test both yourself.

Portrait of Luis Chavez-Mattos
Edited by Luis Chavez-Mattos, Director, Product··RSS

For Engineering and Executives

Comic strip of a young woman looking up in awe at a tall server rack, plugging a network cable into it, and ending on a close-up of the lit port.

Flint or Qwen3.8: which is better for coding?

It depends on where you want to run it. Flint, from Recursive, is an open-weight mixture-of-experts model built for coding agents, so it is the only one of the two you can realistically run on your own machine. Qwen3.8 Max, from Alibaba, is a very large hosted model aimed at harder, long-running engineering work. Use Flint for quick, routine tasks and local privacy. Use Qwen3.8 Max when a job is big and multi-step. No independent head-to-head exists in the sources we reviewed, so test both on your own work.

TL;DR

  • Flint is built to be light. It has 40B total parameters but only about 2.5B active at a time, which is why it targets people with modest hardware.
  • Qwen3.8 Max is a different class of model. It is listed at 2.4 trillion parameters with a 1 million token context window, so it runs on Alibaba's servers, not your laptop.
  • The benchmark wins are mostly vendor-reported. Recursive's published comparison for Flint is against an older Qwen model (Qwen3.6-35B-A3B), not Qwen3.8.
  • Early community reports are mixed. One Reddit poster who promoted Flint later noted that initial reports suggested another model, Ornith-1.5, remains the standard in this size class, possibly due to early software glitches.
  • Some figures circulating online are unverified. Claims like "10x faster" or specific Terminal Bench scores did not appear in the sources we checked, so treat them as unconfirmed.
  • Your own test beats any leaderboard. Run five real tasks from your work on both models and compare.

What are Flint and Qwen3.8?

Flint is an open-weight model from a company called Recursive. "Open-weight" means the model files are published, so you can download and run them yourself. It uses a mixture-of-experts design: the model has 40B parameters in total, but only about 2.5B are used for any given token. It has a 262,144 token context window (the amount of text it can consider at once, roughly a large codebase's worth of files). One listing says it was built by Recursive's automated research system, which they call Eureka Machine. It is described as made for coding agents, meaning tools that read files, run commands and fix their own errors.

Qwen3.8 Max is Alibaba Cloud's flagship in the Qwen family. The AIHubMix listing describes a mixture-of-experts model with 2.4 trillion parameters, a 1 million token context window, and support for text, image and video input. It lists a release date of August 3, 2026. The listing says it focuses on larger engineering projects, long autonomous work and using multiple tools.

One caution on names. You may see "Qwen3.8 27B" in searches. The sources we reviewed cover Qwen3.8 Max, and we could not confirm a 27B model from them. If you mean a smaller Qwen3.8 you downloaded, the advice at the end still applies: test it.

Can you run Flint or Qwen3.8 locally?

Flint, yes. Qwen3.8 Max, no.

"Local" means the model runs on your own computer, so your code never leaves it and you pay no per-token fees. A 2.4 trillion parameter model needs data-center hardware. You reach it through an API (a way for software to send requests to a hosted service), and you pay per use.

Flint is the local candidate. The sources point to community-made GGUF files from Unsloth (GGUF is a common format for running models on consumer machines) and a pull request to add Flint support to llama.cpp, a popular local runner. A pull request is a proposed code change, so support may still have been settling when that post was written.

How big is Flint on disk? The sources do not give a file size, so we won't quote one. Here is the arithmetic you can do yourself. A 4-bit compressed version of a 40B model needs roughly half a byte per parameter, about 20 GB, plus extra memory for the context. That is why figures in the low 20s of GB are plausible, but check the actual file on the download page for the version you pick.

Why the 2.5B active figure matters: all 40B parameters still need to fit in memory, but only a small slice does work on each token. That usually means faster generation than a dense model of the same total size. It does not reduce the memory you need.

Practical reading for your machine:

Your setupWhat to expect
Laptop with 16 GB of memoryFlint's 4-bit files are likely too large. Use a hosted option.
Machine with 32 GB or more (unified memory or GPU)Flint is a realistic candidate. Check the file size first.
Workstation with a large GPUFlint fits comfortably, with room for long context.
Any machineQwen3.8 Max is API only.

Which one is faster?

The sources do not settle this, so be skeptical of any single multiplier you see.

What we can say: the AIHubMix page lists Qwen3.8 Max with a measured time to first token of 35.5 seconds and throughput of 15.9 tokens per second (rolling averages on that service). It shows no equivalent numbers for Flint. So there is no like-for-like speed comparison on that page.

What we can reason: a model with 2.5B active parameters is built to generate quickly, especially on your own hardware with no network wait. A huge reasoning model that thinks before answering will often feel slow on small tasks. That fits the general advice that Flint suits quick, routine edits. But "feels faster" is not "10x faster". Time it yourself.

What do the benchmark numbers actually tell you?

Not as much as the headlines suggest.

  • Terminal Bench measures how well a model completes tasks in a command-line environment, the kind of multi-step work coding agents do. SWE-bench Pro measures fixing real software issues.
  • The Flint claims come from Recursive's own published table: it reportedly beats Qwen3.6-35B-A3B on TerminalBench 2.1 and SWE-bench Pro v2, and ties or betters it elsewhere. The Reddit poster who shared this noted these are the benchmarks the company chose to show.
  • That comparison is against an older Qwen model, not Qwen3.8 Max. A Flint-versus-Qwen3.8 Terminal Bench gap (some summaries cite 73 versus 55) is not something we could verify in the sources, so don't rely on it.
  • Benchmark runs differ in settings, tools and how many attempts a model gets. Vendor tables rarely match what you see on a messy real codebase.

Treat vendor numbers as a reason to try a model, not as proof.

Which should you pick for which job?

Here is a practical split based on what the sources support.

TaskBetter starting point
Rename things, write a small function, explain an errorFlint (fast, cheap, local)
Drafting tests or boilerplateFlint
Working with code you cannot send to a third partyFlint, run locally
Changing many files across a large projectQwen3.8 Max
Long autonomous tasks using several toolsQwen3.8 Max
Tasks that need image or video input (screenshots of a bug)Qwen3.8 Max (Flint is text only)

On cost, AIHubMix lists Flint's free tier at $0 for input and output, and Qwen3.8 Max at $1.69 per million input tokens and $5.07 per million output tokens. A free tier can have limits, so check the terms.

How should you test them yourself?

Pick five tasks from your real work: two easy, two medium, one hard. Give both models the same prompt and judge the results.

Task: [paste the real task, e.g. "Add input validation to this form handler"]
Context: [paste the relevant file(s)]
Constraints: Keep existing function names. Do not add new dependencies.
Output: The changed code and a two-sentence explanation.

Score each run on four things:

  1. Did the code run without edits?
  2. Did it pass your existing tests?
  3. How long did it take from prompt to usable answer?
  4. How many times did you have to correct it?

If Flint passes the easy and medium tasks, use it for those and send only the hard ones to the bigger model. If you run Flint locally and it behaves oddly, update your runner first. Early support glitches were the explanation offered in the community thread.

What does this mean for you?

If you code occasionally (analysts writing scripts, marketers automating a spreadsheet), you do not need to pick a champion. Use a hosted model for anything hard, and try Flint only if you have a capable machine and care about privacy or cost.

If you manage a team with sensitive code, a local model is worth a pilot. Nothing leaves the machine. Run the five-task test and have a developer review the output.

If you buy AI tools, note how little the public comparisons agree. A vendor's table is a starting point for a trial. Ask for a pilot on your own tasks.

Could you skip the model choice and build the tool directly?

For many workers, the real goal is not a coding model but a working tool: a tracker, a request form, a small dashboard. Remy builds and deploys full-stack business apps from a plain-English description, with the database, sign-in and hosting included. You can describe the app, such as a simple log for your model tests with columns for task, model, time and pass or fail, and review the live result. Remy's agents can each use a frontier model you choose, so you are not stuck managing the coding-model decision yourself.

Frequently Asked Questions

5 questions

Is Flint better than Qwen3.8 for coding?

Not in any way the sources can prove. Flint's published wins are against an older Qwen model, and Qwen3.8 Max is a much larger hosted model. Flint is the better fit for local, quick tasks. Qwen3.8 Max is the likelier choice for large, multi-step work.

Can I run Qwen3.8 on my own computer?

Not Qwen3.8 Max. It is listed at 2.4 trillion parameters and is accessed through an API. If a smaller Qwen3.8 variant exists for your hardware, check its own model page for size and requirements.

How much memory does Flint need?

The sources do not state it. As a rough rule, a 4-bit version of a 40B model takes about 20 GB before extra room for context. Check the exact file size on the download page and leave headroom.

Are the Terminal Bench scores for Flint reliable?

Treat them as vendor-reported. The company chose which benchmarks to show, and community reports after release were mixed. Run your own tasks before you commit.

Is Flint free?

AIHubMix lists a free version of Flint at $0 per million input and output tokens. Because the weights are open, you can also run it yourself, which costs only your hardware and electricity.

How we source, write, and correct our articles. Editorial standards

Related articles

Straight to your inbox.
New pieces as they publish. No noise.