$ the-wire · showcase
QWEN3.8 LANDS IN OLLAMA WITH DEVELOPER INSTRUCTION SUPPORT
By RepoJournal · Filed · About Local LLMs · Composed from the cited sources · methodology
16 people shipped this
dhiltgen
@dhiltgen
4 cited
0
@Yunzez
1 cited
fairydreaming
@fairydreaming
1 cited
rchalamala
@rchalamala
1 cited
LucasWilkinson
@LucasWilkinson
1 cited
njhill
@njhill
1 cited
triple-mu
@triple-mu
1 cited
yuan-luo
@yuan-luo
1 cited
Hemanth Battu
@hbattu73
ngxson
@ngxson
lnigam
@lnigam
stefankoncarevic
@stefankoncarevic
connorcarpenter15
@connorcarpenter15
amote-i
@amote-i
b8zhong
@b8zhong
cccccya
@cccccya
Ollama shipped full Qwen3.8 support overnight, adding reasoning-effort semantics and a critical fix for coding agents that send developer instructions before user messages.
The big story across local LLMs this cycle is Qwen3.8 integration. Ollama added the renderer and MLX import support [3], then immediately followed with developer instruction folding [2] to handle OpenAI-compatible coding patterns that were breaking on Qwen3.5. A third commit [4] fixed a runtime message-ordering issue that was turning valid requests into HTTP 500s. On the vision side, Ollama now transcodes WebP images to PNG before sending them to llama-server [1], filling a gap in format support. Over in llama.cpp, the team shipped MiniMax-Text-01 model support [6] and fixed a quadratic performance regression in jinja template parsing [5]. vLLM landed several infrastructure wins: spec decode now taps the correct pre-norm mixture for Kimi K3 DFlash auxiliary states [7], multimodal encoders eliminate blocking GPU-to-CPU syncs [8], and a KV-cache layout refactor pushes quantization knowledge to backends instead of specs [9]. SGLang rebuilt MiniMax-H3 diffusion outputs on demand to save 24.2 GiB of weights [10] and fused MoE activation epilogues to eliminate intermediate buffer round-trips [11].
Action items
- → Upgrade Ollama if you serve Qwen3.8 or accept WebP image payloads ollama/ollama [immediate]
- → Rebuild llama.cpp if you run MiniMax-Text-01 or hit jinja template performance walls ggml-org/llama.cpp [plan]
- → Monitor vLLM spec decode results on K3 if you use DFlash drafting vllm-project/vllm [monitor]
References
- [1] llm: transcode WebP images for llama-server ↗ ollama/ollama
- [2] qwen3.8: support developer instructions ↗ ollama/ollama
- [3] qwen3.8: add renderer and MLX import support ↗ ollama/ollama
- [4] renderers/qwen: tolerate non-leading system messages (#17757) ↗ ollama/ollama
- [5] jinja : fix quadratic cost in gather_string_parts (#27034) ↗ ggml-org/llama.cpp
- [6] model : add support for MiniMaxText01ForCausalLM and MiniMaxM1ForCausalLM (#27018) ↗ ggml-org/llama.cpp
- [7] [Model][Spec Decode] Tap the pre-norm AttnRes mixture as the Kimi K3 DFlash aux state ↗ vllm-project/vllm
- [8] [Perf] Avoid more GPU<->CPU syncs in multimodal encoders ↗ vllm-project/vllm
- [9] [5/N][KV-Cache Layout Refactor] Backend-published KV packing via customize_spec ↗ vllm-project/vllm
- [10] feat(diffusion): rebuild MiniMax-H3 AdaLN outputs on demand ↗ sgl-project/sglang
- [11] [MoE] Fuse swiglu moe up gemm epilogue ↗ sgl-project/sglang
