Build1 publisher3 min readPublished
JetBrains ships a 50/50 weight average of two Qwen checkpoints in Junie Local
The merged 27B finished 37 of 100 internal coding tasks against 34 for the fast checkpoint and 39 for the reasoning one, and it did that while emitting 71% fewer output tokens than the 39, with no post-training in the recipe.
The Engineer · Build desk
What happened
- JetBrains shipped Qwen3.6 with reasoning disabled in the first Junie Local release because the more capable Qwen3.8 needed reasoning enabled, and that made tasks take roughly four times longer.
- For this update it merged the weights of Qwen3.8-27B and Qwen3.6-27B in equal proportions, producing a single 27B model called Qwen3.8-3.6-27B-blend with no additional post-training.
- On a 100-task internal coding benchmark the blend finished 37 tasks, against 34 for Qwen3.6 with reasoning disabled and 39 for Qwen3.8 with reasoning enabled.
- To test whether the blend was simply quitting sooner, JetBrains compared only the 30 tasks both models solved, where the blend used 279K output tokens against Qwen3.8's 935K.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- capability Reproducing this needs no training run. Any team holding both checkpoints can average the tensors and score the result on its own task set, because the two models share an architecture.
- contradiction The visual benchmarks cut against the token story: there the blend reasoned longer and spent more than Qwen3.8, so the saving belongs to this coding workload and has to be re-measured on anything else.
- constraint No one outside JetBrains can check whether the ordering survives a different mix of tasks, because the evaluation that produced the 37 is the company's own.
- decision JetBrains made the merge its shipping model while distillation and reinforcement learning experiments continue, so those methods now have to beat a baseline that cost one averaging pass to produce.
A weight average is only defined when the two checkpoints have the same shape. Qwen3.8-27B is built on Qwen3.6-27B and shares its architecture, so JetBrains averaged the two parameter sets in equal proportion and got a single 27B model out, with no additional post-training [1]. Averaging two sets of numbers is not where I would have gone looking for a capability gain. JetBrains lists distillation, reinforcement learning and more elaborate merging methods among the routes it is still testing [13].
The token figure is the one most likely to be an artifact, and JetBrains went after it. A model that gives up sooner emits fewer tokens without being more efficient. So the comparison was restricted to the 30 tasks both Qwen3.8 and the blend solved [5]. One long trace can skew a mean, so the per-task count is the check that holds: the blend was cheaper on 29 of the 30, at 279K output tokens against 935K [5][6], about 9,300 per solved task against 31,200 [17].
The overlap number carries a second result. Qwen3.8 solved 39 and the blend 37, with 30 in common, so the blend finished 7 tasks the reasoning model failed and failed 9 that it finished [18].
On LiveCodeBench the ordering held over four runs: 85.47% for the blend against 83.29% for Qwen3.8, at similar output cost [7]. Qwen3.6 averaged 67.87% there and spent about 24.1 million output tokens a pass against roughly 6.14 million for the blend [8], a factor of about 3.9 [19]. The visual benchmarks went the other way. There the blend spent more tokens than Qwen3.8 on identical questions, images and generation settings, and JetBrains attributes the extra almost entirely to the blend spending more time on reasoning [9].
Chain-of-thought length also varied significantly across four identical runs, and JetBrains calculates that taking the shorter correct trace would have cut token use by 24.5% [10]. The company says that suggests shorter successful paths exist and that it could potentially teach the model to take those paths with zero performance loss [11].
For the coding numbers to transfer, a team's tasks have to resemble the 100 in an evaluation JetBrains describes as internal [2], and those results are given as single solve counts while the LiveCodeBench figures are averages of four runs [15]. Three solves out of 100 is a thin margin to carry into a different task mix, and the blend is still two behind Qwen3.8 [20]. The same update adds experimental NVIDIA support on Windows [14].
JetBrains says the blend can still overthink when it struggles to find a solution, and the workaround it recommends is manual [12]. "If Junie keeps revisiting the same approach without new evidence or useful tool results, we recommend interrupting it and restarting it with a narrower goal," the company wrote [16].
What to watch
- Publication of the blend weights or the internal task list, which would let outside teams reproduce the 37-versus-34 result.
- Whether the distillation and reinforcement learning experiments JetBrains says are continuing beat a merge that cost no training run.
- Whether the 24.5% short-trace headroom turns into a shipped token reduction, by trace selection or by further training.