Which models can be cast into Fanshi silicon?
Customers ask us most about the newest Qwen releases. Qwen3.6 and Qwen3.8 keep the Qwen3.5 hybrid layout, which Fanshi inference cores already map, so a newer model in the same family is a new chip build on existing cores rather than a new architecture.

Model families and how they map
This is architecture-level fit. Parameter size, context length and performance targets are assessed per project.
Model family
Architecture
Fanshi path
Status
Qwen3.5
Architecture
Hybrid: Gated DeltaNet linear-attention layers + gated full-attention layers
Fanshi path
Forge (0.8B) · Trident (up to 35B) · FS-Linear + FS-Attention
Status
Core model of the first-generation chips
Qwen3.6 / Qwen3.8
Architecture
Same hybrid layout as Qwen3.5
Fanshi path
FS-Linear + FS-Attention · custom build
Status
Maps onto existing cores; assessed per project
Qwen3
Architecture
Full attention (GQA)
Fanshi path
FS-Attention
Status
Supported via IP licensing
SmolLM2
Architecture
Compact full-attention language models
Fanshi path
FS-Attention
Status
Supported via IP licensing
SmolVLM · SegFormer
Architecture
Vision-language understanding and semantic segmentation
Fanshi path
Argus · FS-Fusion
Status
Core models of the first-generation chips
Sparse MoE models
Architecture
Dynamically activated sparse experts
Fanshi path
FS-MoE
Status
Assessed per project
DeltaNet-style linear attention
Architecture
Fixed-size recurrent state, streaming
Fanshi path
FS-Linear
Status
Supported via IP licensing
Figures on this page are current product definitions and design targets. Final specifications, test conditions, availability and delivery versions are confirmed in project documents agreed by both parties.
Why the Qwen hybrid layout suits silicon
From Qwen3.5 onwards, Qwen models stack their layers in groups of four: three Gated DeltaNet linear-attention layers followed by one gated full-attention layer. The Qwen3.5, Qwen3.6 and Qwen3.8 27B models all use the same 16 × (3 + 1) layout.
A linear-attention layer carries a fixed-size state from token to token instead of a growing KV cache, so only one layer in four keeps a cache that grows with context. On edge silicon, where on-chip memory and bandwidth are the scarce resources, that is often the difference between a model that fits and one that does not.
Fanshi’s FS-Linear core handles the Gated DeltaNet layers and FS-Attention handles the gated full-attention layers. Because the layout has not changed across Qwen3.5, 3.6 and 3.8, the same cores map all three generations.
When a newer model ships
What changes depends on where the weights live.
MaskROM
New weight masks
Re-make the mask layers that hold the weights. With the same architecture, the core design, SDK and host integration carry over.
SRAM CIM
Reload
Load the new weights at boot.
NMC
Update memory
Update the weights in external memory; the model can keep iterating.
What we need for a feasibility check
- Company and your role
- Target model(s) and parameter size
- Context length (typical / maximum)
- Usage pattern (single-shot, multi-turn, agentic)
- Throughput and latency goals
- Power budget and form factor
- Expected volume and timeline
Questions about models
Do you support Qwen3.6, Qwen3.8 and other new models?
Yes. Qwen3.6 and Qwen3.8 keep the Qwen3.5 hybrid layout: Gated DeltaNet linear-attention layers interleaved with gated full-attention layers. Fanshi’s FS-Linear and FS-Attention cores already cover both layer types, so a newer model in the same family is a new chip build on existing cores, not a new architecture. Parameter size, context length and performance targets are assessed per project.
What happens when a newer model is released after the chip is made?
On the MaskROM route, moving to a new model means re-making the mask layers that hold the weights. If the new model shares the same architecture, the core design, SDK and host integration carry over unchanged. On SRAM compute-in-memory you reload the weights at boot; on near-memory compute you update the weights in external memory.
Which models do Fanshi chips support?
Forge targets Qwen 3.5 0.8B; Trident supports up to Qwen 3.5 35B; Argus runs SmolVLM, SegFormer and similar vision and multimodal models. The IP family covers Qwen3 / 3.5 (0.6B to 9B), SmolLM2, sparse MoE models and DeltaNet-style linear attention. See the supported models page for the full list.
Further reading
Qwen3.5 to Qwen3.8: Why Hybrid Attention Suits Edge SiliconQwen3.5, 3.6 and 3.8 interleave Gated DeltaNet linear attention with gated full attention at 3:1. Why that layout keeps memory in check on edge AI silicon.
How to Choose an Edge LLM Inference Chip: 7 QuestionsChoosing an edge LLM inference chip comes down to model fit, power budget and how the model updates, not peak TOPS. Seven questions to answer, plus a checklist.
MaskROM Compute-in-Memory: What Casting Weights in Silicon MeansMaskROM compute-in-memory writes model weights into a via array at the mask stage, so nothing moves at runtime. How it works, why it costs least, its limits.
Start from your model
Send the target model, context length, usage pattern and power goal. We reply with a route and an evaluation plan.
