ClovisChat - Qwen3.8-Flash-Next
ClovisChat will move to Qwen3.8-Flash-Next, replacing Qwen3.6-35B as its conversation and vision model. The ClovisChat alias will remain unchanged: this transition only affects the underlying model.
This transition aims to improve everyday assistance: understanding a request, reasoning through a problem, and supporting work across multiple steps.
Key figures
ClovisChat currently uses Qwen3.6-35B. The charts compare the official scores for Qwen3.6-35B with those for Qwen3.8-Flash-Next.
The figures come from the official model cards for Qwen3.6-35B-A3B and Qwen3.8-Flash-Next.
GPQA Diamond — Scientific reasoning — +5.7 points
HLE — Multidisciplinary reasoning — +14.5 points
LiveCodeBench v6 — Solving programming problems — +11.5 points
RealWorldQA — Real-world visual understanding — +3.2 points
These scores are technical indicators. They do not directly measure the quality of a conversation in French or ClovisChat’s response time.
Why some gains are much larger than others
The reported gap is larger on HLE than on RealWorldQA. These tests assess different skills: multidisciplinary reasoning and image understanding. This evolution therefore cannot be summed up as a single percentage improvement.
For ClovisChat, the priority use cases to evaluate are:
Comparison with other models
The table published alongside Flash-Next also shows how the model compares with other reference models:
| Benchmark | Qwen3.8-Flash-Next | Qwen3.7-Plus | DeepSeek-V4-Flash-0731 | Claude Opus 4.6 (Max) |
|---|---|---|---|---|
| GPQA Diamond | 91.7 | 90.3 | 90.8 | 91.3 |
| HLE | 35.9 | 34.7 | 33.8 | 40.0 |
| LiveCodeBench v6 | 91.9 | 89.6 | 90.6 | 88.8 |
Flash-Next leads this selection on two benchmarks; Opus retains the advantage on HLE. This table is not an overall ranking. Source: official Qwen evaluations.
Practical improvements in software development
Qwen3.6-35B-A3B scores 49.5 on SWE-bench Pro, using an in-house agent and a 200K-token context window. Source: Qwen3.6-35B-A3B.
Flash-Next reaches 62.5, using Claude Code and a 256K-token context window. Because the evaluation setup changed, the gap does not measure improvements in the model alone. Source: Qwen3.8-Flash-Next evaluation setup.
In a coding conversation, the intended benefits include:
A new architecture designed for efficiency
Flash-Next previews the architecture of Qwen4. Its main model has 125 billion parameters, with 6 billion activated per token, plus 51 billion N-gram embedding parameters. These figures describe different components, so 125 billion does not account for all the weights. Source: official Qwen repository.
In ClovisChat, the maximum context window is set to 262,144 tokens, including both input and output. A response can contain up to 32K tokens; when this full output capacity is reserved, approximately 230K tokens remain available for the prompt, conversation history, and provided documents.
The attention mechanism combines Gated DeltaNet with Qwen Sparse Attention, which selects relevant context blocks. Gated residual connections improve information flow between layers, while N-gram embeddings add capacity with little extra computation. The goal is to use resources more efficiently, particularly for long sequences. Source: Qwen technical report.
What Qwen3.8-Flash-Next does not solve
A higher score does not guarantee that any given answer is correct. Factual answers, suggested code, and actions performed with tools must remain verifiable.
For professional use, the same best practices still apply:
- check critical information against its sources;
- test code before integrating it;
- limit tool permissions to what is needed;
- retain human approval for critical operations.