Two new coder-model benchmark claims are circulating without a single independent number behind them: Recursive's Flint 40B claiming 89.7% on SWE-bench Pro v2, and IQuest's Q1 320B claiming 83.2% on Terminal-Bench 2.1. The story is less the scores than the verification gap around them.
Key facts
- Recursive Flint 40B/2.5B MoE claims 89.7% on SWE-bench Pro v2 under Apache 2.0 — lab self-report, unverified.
- IQuest Q1 320B/15B MoE claims 83.2% Terminal-Bench 2.1 and 57.5 SWE-bench Pro — lab self-report, unverified.
- DeAI covered Flint's open-weights release on October 22; the new element is the 89.7% claim.
What's driving the conversation
AI Weekly's open-source alert feed surfaced both claims this week, and they are spreading through the usual chain: aggregator post, quote posts, leaderboard screenshots. Flint's 89.7% on SWE-bench Pro v2, if it held under independent testing, would put an Apache-2.0-licensed 40B model at the top of an agentic-coding leaderboard. Q1's 83.2% Terminal-Bench 2.1 claim would do the same for the 320B class. Neither claim has a public eval config attached, and the alert snippets do not date the Q1 figure.
The substance
Both numbers are provider self-reports. DeAI's October 22 Flint coverage reported the release and its licensing without the benchmark claim, which only now has a specific figure attached. What would verify: an independent SWE-bench Pro v2 or Terminal-Bench 2.1 run, a public eval harness configuration, or a grade from an independent checking harness — not another repost. Release dates for both models are not visible in the alert snippets either, so even the timeline is soft. Until any of that lands, these are claims, and they should be planned around as claims.
Why builders are watching
Benchmark self-reports are how new coder models get adopted before they are tested, and the pattern ends one of two ways: an independent run confirms and the model jumps the queue, or the number quietly never reproduces and the repo keeps the star count anyway. For teams benchmarking the dense-32B class against Qwen 3.8 and Nanbeige 4.3-33B, the practical move is to wait for a reproduction before adding either model to an eval matrix. If you test Flint yourself, publish the harness config — that is the only thing that moves these numbers from claim to fact.
Questions
- What are the Recursive Flint and IQuest Q1 benchmark claims?
- Aggregated alerts this week list Recursive's Flint, a 40B/2.5B MoE coder claiming 89.7% on SWE-bench Pro v2 under Apache 2.0, and IQuest's Q1, a 320B/15B MoE agentic CLI model claiming 83.2% on Terminal-Bench 2.1 and 57.5 on SWE-bench Pro. Both figures are lab self-reports with no published independent run.
- Was Recursive Flint already covered before?
- Yes. DeAI covered Flint's open-weights release on October 22. The new element this cycle is the specific 89.7% SWE-bench Pro v2 claim, which was not part of the earlier post. The Q1 claim's recency is also unconfirmed.
Sources
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.