Qwen3-8B
An 8-billion-parameter autoregressive language model from Alibaba's Qwen3 series, used as an autoregressive planner baseline in comparative experiments.
Timeline (last 90 days)
Qwen2.5-3B was excluded because its license is marked as other, and Qwen3-8B was excluded because it exceeds memory; DeepSeek-R1-Distill-Qwen-1.5B and OLMo 2 1B, although they meet the license requirements, cannot be loaded on this vLLM image.
On Qwen3-8B, after the second round of the loop selected 3 new skills, model performance improved from 64.3% to 66.3%.
In this test configuration, the 8B model handles simple tasks (short translation, extraction, Q&A, arithmetic), the 32B model handles medium tasks (SQL, algorithm implementation, debugging), and DeepSeek handles complex tasks (concurrent code, distributed architecture design).
Compared to the autoregressive model, the diffusion planner reduced the average latency of plan generation by 39%–46%.
On the Natural Plan benchmark without task-specific training, the diffusion model's plan repair success rate was 53.7%, while the autoregressive model's was 27.0%.
The research team compared the diffusion-based planner DreamReasoner-8B at the 8B scale with the autoregressive planner Qwen3-8B.
After applying WavePrune to Qwen3-8B, the HELMET score improved from 35.7 to 40.0.
CLM's 8-billion-parameter backbone Qwen3-8B is frozen and serves only as a reader; what is actually trained are only the small projection heads on each side with approximately 20 million parameters each, and a complete pretraining run takes about one hour on a single RTX.
When using Qwen3-8B for constrained decoding, latency climbs above 4 seconds as the number of candidates increases.
The RL-Kernel team ran a 200-step Qwen3-8B GRPO training task on 8 AMD MI300X cards, with results showing zero logprob mismatch between the Megatron training side and the vLLM rollout side.
13 more timeline events
All Facts (20)
Re-quantizing INT4 weights to NF4 subjects each weight to two rounds of rounding, and official measurements show this increases the perplexity of Qwen3-8B by 1.6%
70%UnverifiedQwen2.5-3B was excluded because its license is marked as other, and Qwen3-8B was excluded because it exceeds memory; DeepSeek-R1-Distill-Qwen-1.5B and OLMo 2 1B, although they meet the license requirements, cannot be loaded on this vLLM image.
50%UnverifiedOn Qwen3-8B, after the second round of the loop selected 3 new skills, model performance improved from 64.3% to 66.3%.
50%UnverifiedIn this test configuration, the 8B model handles simple tasks (short translation, extraction, Q&A, arithmetic), the 32B model handles medium tasks (SQL, algorithm implementation, debugging), and DeepSeek handles complex tasks (concurrent code, distributed architecture design).
50%UnverifiedCompared to the autoregressive model, the diffusion planner reduced the average latency of plan generation by 39%–46%.
50%UnverifiedOn the Natural Plan benchmark without task-specific training, the diffusion model's plan repair success rate was 53.7%, while the autoregressive model's was 27.0%.
50%UnverifiedThe research team compared the diffusion-based planner DreamReasoner-8B at the 8B scale with the autoregressive planner Qwen3-8B.
50%UnverifiedAfter applying WavePrune to Qwen3-8B, the HELMET score improved from 35.7 to 40.0.
50%UnverifiedCLM's 8-billion-parameter backbone Qwen3-8B is frozen and serves only as a reader; what is actually trained are only the small projection heads on each side with approximately 20 million parameters each, and a complete pretraining run takes about one hour on a single RTX.
50%UnverifiedWhen using Qwen3-8B for constrained decoding, latency climbs above 4 seconds as the number of candidates increases.
50%UnverifiedThe RL-Kernel team ran a 200-step Qwen3-8B GRPO training task on 8 AMD MI300X cards, with results showing zero logprob mismatch between the Megatron training side and the vLLM rollout side.
50%UnverifiedCLM is a set of lightweight projection heads built on the Qwen3-8B backbone, totaling approximately 75MB in size, and is positioned as an open-weight, locally deployable alternative to the closed-source model Jev.
50%UnverifiedThe research team validated LEGO using Qwen3-8B as the backbone model.
50%UnverifiedThe 8B-scale model using the LEGO framework performs comparably to larger-scale models on professional legal reasoning tasks.
50%UnverifiedThis rotation-decay geometric phenomenon has been reproduced on three models: Llama-3.1-8B, Qwen3-8B, and Gemma-4-E4B.
50%UnverifiedThe recommended division of labor is for the Hermes brain to handle task understanding, planning, and scheduling; the local model to handle specific execution tasks; and a high-spec model to handle long contexts and complex reasoning.
50%UnverifiedLocal models are not well-suited for complex multi-step reasoning or long-chain inference.
50%UnverifiedThe Qwen3-8B distilled model is capable of handling execution-type tasks with clear boundaries, such as content extraction, text organization, classification, format conversion, simple data processing, and basic coding.
50%UnverifiedIn the study, the governed mode uses the Qwen3-8B model solely for interpreting user intent, with analytical procedure execution delegated to a deterministic policy engine.
50%UnverifiedOsprey achieves a 16.1% improvement in mean acceptance length on Qwen3-8B.
50%