Flint v1.0 vs Qwen3.8 27B: The Model Recursive Benchmarked Itself Against

Recursive published one sentence in its Recursive Flint v1.0 launch material that most coverage skipped: on the company's own six-test average, only Qwen3.8 27B scored higher than Flint. That is an unusual thing for a vendor to print on launch day. It is also the most useful sentence in the release, because it tells you which comparison to go and check. Qwen3.8 27B is an open-weight dense model, released on 14 August 2026 under Apache 2.0, and it is the only model in Recursive's comparison set that beat Flint on the aggregate the vendor itself constructed. So the interesting question is not whether Flint is good. It is where, exactly, Qwen3.8 27B is better, and whether the gap survives scrutiny of the harnesses involved.

Read the class claim before you read the score

Flint's headline claim is carefully bounded: it leads mixture-of-experts models under 50 billion total parameters on SWE-bench Pro v2, Terminal-Bench 2.1 and SWE-bench Verified. Qwen3.8 27B is not a mixture-of-experts model. It is dense — 27 billion parameters, all of them active for every token, 28 billion counting the vision encoder, 64 layers at a hidden size of 5,120. So it sits outside Flint's comparison class by construction, and Recursive was free to leave it out of the claim entirely.

It is worth being precise about what "dense" costs and buys. Every one of Qwen's 27 billion parameters participates in every token, which is why it needs about 55.6 GB in BF16 and a real GPU, and why its activation footprint is roughly eleven times Flint's. It is also why a dense 27B can outperform a 2.5B-active sparse model on accuracy: it is doing more arithmetic per token, and it is charging you for it.

Terminal-Bench 2.1: the eighteen-point gap

This is the cleanest comparison available, because both labs report the same benchmark version, and the gap is large enough that harness noise is unlikely to explain all of it.

• Flint — 54.9% on Terminal-Bench 2.1, run on the Terminus 2 harness with internet enabled, but with any solution obtained through web retrieval counted as a failure.

• Qwen3.8 27B — 73.0% on Terminal-Bench 2.1, per Alibaba's model card.

That is an 18.1-point gap in Qwen's favour, on the benchmark that most closely resembles agentic terminal work — which is precisely the job both models are sold for. Two caveats cut in opposite directions and both need stating.

The first favours Flint. Recursive applied a handicap its peers may not have applied: if the agent found the answer on the web, Recursive scored it as wrong. A model allowed to search and cite would score higher on the same harness. How much higher is unknown, and Recursive does not publish the un-handicapped figure.

The second cuts against any clean reading. A third party, PrismML, scored Qwen3.8 27B at 69.7 on Terminal-Bench 2.1 rather than Alibaba's 73.0. A 3.3-point disagreement on the same model and benchmark is a useful calibration: it tells you that gaps of a few points in this table are noise, and that an 18-point gap is probably not.

Even taking the lower Qwen figure, 69.7 against 54.9 leaves 14.8 points. The direction is not in doubt.

LiveCodeBench v6: a second gap, on a cleaner benchmark

Competitive programming is the least harness-sensitive of the benchmarks both labs report, because the model gets a problem and produces code and a judge runs it. There is no agent loop to differ over.

• Flint — 80.3% on LiveCodeBench v6.

• Qwen3.8 27B — 90.3% on LiveCodeBench v6.

That is a 10-point gap, and it is the one I would weight most heavily, precisely because it is the hardest to explain away. Flint's 80.3 is a respectable number for a model that activates 2.5 billion parameters. Qwen's 90.3 is a strong number for a model that activates 27 billion. Neither is surprising. The point is that the efficiency advantage Flint is selling does not come free, and this benchmark is where the price is visible.

The benchmark that looks like a Flint win and is not

Here is where most comparison pages go wrong, and it is worth slowing down for.

Flint reports 89.7% on SWE-bench Pro v2. Qwen3.8 27B reports 61.7% on SWE-bench Pro. Read quickly, that looks like a 28-point Flint victory. It is not a comparison at all. The "v2" is a different benchmark from the one Alibaba reports, with a different task set and a different evaluation protocol, and the two numbers cannot be subtracted. Any page that puts them in adjacent columns and declares a winner is telling you something about its author, not about the models.

The same trap sits one row down. Flint reports 74.2% on SWE-bench Verified. Qwen3.8 27B does not publish a SWE-bench Verified score — Alibaba's card reports SWE-bench Pro, QwenSWEBench at 79.0, DeepSWE 1.1 at 42.2, Terminal-Bench 2.1 at 73.0, LiveCodeBench v6 at 90.3 and NL2Repo-Bench at 42.3. There is no Verified figure to compare against. Numbers circulating for Qwen3.8 27B on Verified are either inherited from an earlier model in the family or taken from a secondary source that conflicts with the model card on several other benchmarks at once. Treat them as unverified.

One further qualification on Alibaba's side: its SWE-bench Pro and QwenSWEBench runs used an external agent harness rather than a bare model, so even comparisons against another lab's SWE-bench Pro figure are shaky. This is not unique to Alibaba. It is the state of agentic coding benchmarks in general, and it is why the Terminal-Bench and LiveCodeBench rows carry more weight here than the SWE-bench rows do.

Where Flint wins, and it is not close

Accuracy is not the only axis, and on the others the ordering reverses hard.

• Memory — Flint runs in roughly 24 GB, with a 22.86 GB NVFP4 build on disk. Qwen3.8 27B needs about 55.6 GB in BF16. One fits a high-end laptop or a single consumer card; the other needs a real GPU allocation.

• Active compute per token — Flint activates about 2.5 billion parameters. Qwen activates all 27 billion. That ratio is the reason the speed gap exists and the reason Flint is cheaper to serve at volume.

• Speed — Recursive claims 125-plus tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64 GB RTX Spark at short context. A dense 27B on the same laptop is not in that range. Qwen3.8 27B is also reported to be slow locally and unusually token-hungry — Artificial Analysis measured roughly 160 million output tokens across its index testing, which is a cost multiplier in any agent loop.

• Context — both document 262,144 tokens natively, and Qwen extends to 1,000,000 with YaRN scaling. Flint's announcement cites up to 512K. On paper this is a wash at the native window; in practice Flint's throughput at long context is the weaker story, with independent testing reporting a fall to roughly 9 tokens per second near 115,000 tokens.

• Multimodality — Qwen3.8 27B accepts images and video as well as text, with a hybrid attention layout of 48 Gated DeltaNet linear-attention layers to 16 full-attention layers at a 3:1 ratio, which is what makes a 262K window affordable to serve. Flint is text only.

What you are actually choosing

• If maximum coding accuracy is the goal and you have the hardware, Qwen3.8 27B is the stronger model on every benchmark where the two can be compared cleanly. That is also what Recursive's own six-test average says.

• If the model has to run on a laptop, stay offline, keep source code on the machine, or serve enough volume that active parameters dominate the bill, Flint wins on the axis that decides the deployment. A 2.5B-active model at 125 tokens per second is a different product from a 27B dense model that needs a GPU node.

• If you want both, they compose well: the local model for routine and sensitive work, the hosted model for the hard tasks. That split is the one most teams land on, and it is where the routing layer stops being a detail.

Qwen3.8 27B is the one model in this comparison that OrcaRouter serves directly — the identifier is qwen/qwen3.8-27b in our live model list, and it accepts text, images and video with the full 262,144-position window. Because provider list price is passed through with no markup added, a price change on Alibaba's side is live on our side the same day rather than at the next renewal. OrcaRouter does not serve Flint; that one is a download. So the hybrid above is genuinely hybrid: one path to a hosted frontier-class model, and a local file for everything that must not leave the machine.

The experiment nobody has run

Every number in this article came from a lab measuring its own model, with one exception — PrismML's 3.3-point disagreement with Alibaba on Terminal-Bench, which is the only hint available about how much these figures move when someone else holds the stopwatch.

What would settle this is straightforward and has not happened: both models, one harness, one task set, one machine, published traces. Until then the defensible position is the narrow one. On the benchmarks both labs report honestly, Qwen3.8 27B is the more accurate coder. On the deployment axes that decide whether a model can run at all, Flint is the more efficient one. Recursive said so first, in its own launch post, and that is the most trustworthy thing in the entire release.