Qwen3.8
Alibaba
Tidslinje
- · Kildelinks · GitHub search · Alibaba · commits · onprem-ai/deployment-catalog · @tomas-polach
- · Kildelinks · GitHub search · Alibaba · commits · finereli/pyramid-mcp · @finereli
- · Kildelinks · GitHub search · Alibaba · commits · fdietze/pi-infinite-context · @fdietze
- · Alex' indlæg · modelfamilie · GitHub search · Alibaba · commits · agentgateway/agentgateway · @github-actions[bot]
Uddrag
json.providers.nebius.models["Qwen/Qwen3.8-27B"].name = "Qwen3.8 27B";
- · Kildelinks · GitHub search · Alibaba · commits · aabhay007/Personal-AI-Core · @aabhay007
- · Kildelinks · GitHub search · Alibaba · commits · QuantumKuba/Diffactly · @QuantumKuba
- · Kildelinks · GitHub search · Alibaba · commits · Hallo-ich-bin-PeppaPig/Singularity · @Hallo-ich-bin-PeppaPig
- · Kildelinks · GitHub search · Alibaba · commits · Azure/azureml-assets · @hanbitmyths
Uddrag
Add MAI, Qwen3.8 and Nemotron 3.5 Foundry Local coding models (#5452)
- · Alex' indlæg · modelfamilie · GitHub search · Alibaba · pulls · xadupre/devdash-agents · @xadupre
Uddrag
- Real Qwen3.8 27B on H200-101, physical CUDA 7: streaming generation and a follow-up turn succeeded. Used a temporary configuration with 16 KV blocks and
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] qwen38_flash package entry: register() and READMEs
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] MTP draft model
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] KV offload pool regions
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Runner input adapter
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Hyper-connection layers
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] GDN layer input preparation
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Pooled KV cache materialization
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] PD transfer: identity, manifest and Raiden connector
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Config hook, geometry checks and pooled cache layout
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] PLE feature: tables, host lookup, storage, offload and pooled PLE
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Sequence-parallel prefill backbone
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] QSA attention layer
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Model-owned GDN op and layer (no register_oot)
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Per-group Mamba index accessors
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Mamba tail fold: export prefill checkpoints from inside the step
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] In-place rows update op
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] PLE Torch-to-JAX ops
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] QSA Torch-to-JAX ops
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Torch bridge for the QSA kernels
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] GDN input preparation kernel
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Unified KV pool kernels
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] PLE lookup and gate kernels
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Hyper-connection kernels
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Pooled QSA kernel over the unified KV pool
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] QSA KV write kernels (fused, main, compressed)
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] QSA union-walk kernels (plan, dense, SparseCore gather, prefill)
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] Qwen3.8-Flash model and registration hooks
- · Kildelinks · GitHub search · Alibaba · pulls · vllm-project/vllm-torchtpu · @DRZDONG
Uddrag
[Model][Qwen3.8] QSA main-KV decode context parallelism over rank pairs (DCP-2) kernels
- · Kildelinks · GitHub search · Alibaba · pulls · vjmichael/Lu-AI-GUI · @vjmichael
Uddrag
- Physical interactive owner Windows/Radeon Qwen3.8, Vulkan VRAM/TFFB TTFT p50/p95, and actual local installed MMProj; hosted CPU small models do not substitute.
- · Kildelinks · GitHub search · Alibaba · pulls · vitessio/vitess · @jdoupe
- · Kildelinks · GitHub search · Alibaba · pulls · nearai/cvm-compose-files · @PierreLeGuen
- · Kildelinks · GitHub search · Alibaba · pulls · michsindlinger/specwright · @michsindlinger
- · Alex' indlæg · modelfamilie · GitHub search · Alibaba · pulls · SceneWorks/SceneWorks · @michaeltrefry
Uddrag
SW-7119: Qwen3.8 27B describer model and tier in core, worker, API and catalog
- · Alex' indlæg · modelfamilie · GitHub search · Alibaba · pulls · NVIDIA/hyperflare-inference · @Raydius
Uddrag
The pinned Qwen3.8 Flash Next checkpoint trains a 262,144-token context (theta 10M, interleaved MRoPE 11/11/10). This adds a declared serving variant of the same weights
- · Kildelinks · GitHub search · Alibaba · pulls · FlashML-org/FreeToken · @zihaomu
- · Alex' indlæg · modelfamilie · GitHub search · Alibaba · pulls · BuQQzz/SHIELD2.0 · @BuQQzz
Uddrag
Makes Qwen3.8 27B (IQ1_M) a working coding model on an 8 GB card (RTX 2070 SUPER) through SHIELD's own llama-server.
- · Kildelinks · GitHub search · Alibaba · pulls · Azure/azureml-assets · @hanbitmyths
Uddrag
Add MAI, Qwen3.8 and Nemotron 3.5 Foundry Local coding models
- · Kildelinks · GitHub search · Alibaba · pulls · 0xburden/modelship · @0xburden
Uddrag
- Reviewed immutable catalog of 25 standalone Qwen3.8 GGUF quantizations; wizard displays sizes, numeric/letter/full-value selection and a visible UD-Q4_K_M default. No hardware-driven quant substitution; flags
- · Kildelinks · GitHub · QwenLM/qwen-code · PR
Uddrag
- The Standard preset has the same qwen3.8 effort and GLM thinking gaps and will be aligned in a follow-up
- · Alex' indlæg · modelfamilie · Alex Getman
Uddrag
🚨 Qwen3.8 Omni Flash is out It can see, hear and understand video, then actually act on it Give it hours of footage and it can find the right moments, translate, edit or turn it into notes. Give it a meeting and it can pull out tasks and start working on them 1M context, API from $0.43
- · Alex' indlæg · modelfamilie · Alex Getman
Uddrag
🚨 QWEN3.8 FLASH NEXT IS OUT The leaked specs were real: • 125B multimodal MoE • Only 6B active • 262K native context, up to 1M Qwen says it beats Claude Opus 4.6 across coding, agentic tasks, tool use and several multimodal benchmarks.
- · Alex' indlæg · modelfamilie · Alex Getman
Uddrag
🚨 QWEN3.8 FLASH NEXT DROPS TODAY A ModelScope description that was quickly taken down reportedly revealed: • 125B main parameters • 51B N gram embeddings • Only 6B active per token • Multimodal MoE architecture The model is also an official preview of the upcoming Qwen4 architecture.
- · Alex' indlæg · Alex Getman
Uddrag
⚡️ Qwen 3.8 27B beat GPT 5.6 Terra This happened in the AA Agentic Index, a benchmark that tests how well models handle agentic tasks. And you can run Qwen 3.8 27B locally on a single RTX 3090.
- · Alex' indlæg · Alex Getman
Uddrag
⚡️ Alibaba released Qwen3.8’s weights A model at Claude Opus 4.6 level, a flagship just six months ago, can now run on a single RTX 4090. Qwen3.8 27B uses around 17 GB of VRAM when quantized and should deliver ~50 tok/s.
- · Alex' indlæg · Alex Getman
Uddrag
Alibaba just teased Qwen3.8 with 2.4T parameters This is the new Qwen flagship, with open weights coming soon. Alibaba ranks it right below Claude Fable 5. Public benchmarks are still missing, so nobody really knows how close it actually is to Fable, or how it stacks up against GPT-5.6 Sol and Kimi K3. Pricing is the fun part. Kimi K3 costs $3/$15, so Qwen3.8 will probably land in the same range. The preview is already live in Qoder. It uses 10x fewer credits during the day and 50x fewer at night, so you can basically test it for free.
Specifikationer
Radar · Qwen3.8 · Ukendt lanceringsdato
Spørgsmål
Hvornår blev Qwen3.8 lanceret?
Ukendt lanceringsdato