BlackRiver AI / Flagship local model
A BlackRiver post-trained sparse MoE built from Qwen3.6-35B-A3B for the work we actually care about: coding, agents, long-horizon reasoning, tool use and local multimodal systems. 35B total parameters. ~3B active. 262K native context. Vision. Native MTP.
Qwen3.6-35B-A3B is already an unusually strong small-active MoE. We did not build QwiVer because the base was weak. We built it because good foundations can still be pushed into a better working model.
A powerful 35B sparse MoE with vision, long context, thinking mode and strong agentic coding capability. The architecture is the reason QwiVer exists.
Same exceptional sparse architecture. Different behavior. QwiVer is the version BlackRiver chooses for day-to-day coding and agentic work because it is more useful in the workflows that actually consume our time.
That is the shortest description after living with the model every day. Not because a 3B-active local MoE suddenly has frontier-scale coverage, but because QwiVer approaches work with the kind of deliberate, structured, tool-minded persistence we normally associate with much larger coding models. It stops feeling like “a surprisingly good local model” and starts feeling like an actual collaborator.
Behavioral comparison is BlackRiver's subjective daily-use shorthand. It is not a claim of benchmark equivalence, affiliation or identical capability to GPT-5.6 Sol or Fable 5.
The published GGUFs come from the merged BlackRiver model, not an untouched Qwen checkpoint with a new filename. The Phase 10.2 adapter was trained, selected, merged into the BF16 language model and then converted through an MTP-aware release pipeline.
1,531 examples · 3,999,924 training tokens · 383 optimizer steps. Small enough to stay surgical, deliberate enough to materially change the model we use.
The selected Phase 10.2 LoRA was merged directly into the frozen BF16 language-model base before GGUF conversion. The release is self-contained; no runtime adapter is required.
QwiVer preserves the architecture that makes Qwen3.6 special: sparse routing, vision, 262K native context and MTP tensors for compatible speculative decoding.
QwiVer inherits Qwen3.6's sparse MoE design: a 35B parameter pool with only ~3B parameters activated per token. That is the entire attraction — large-model capacity without dense-model economics.
The router selects a tiny active path through a much larger expert pool. QwiVer keeps that efficiency while changing the post-trained behavior riding on top of it.
Four MTP-complete GGUF editions cover memory-constrained systems through near-lossless quantized deployment. The Q4_K_XL release is the BlackRiver default when there is enough memory to run it comfortably.
For systems where fitting the model matters more than preserving every last bit of fidelity.
QwiVer3.6-35B-A3B-UD-Q2_K_XL.ggufA stronger low-memory compromise for local machines that cannot comfortably stretch to Q4.
QwiVer3.6-35B-A3B-UD-Q3_K_XL.ggufThe release we would start with: practical footprint, strong fidelity and room left for meaningful context.
QwiVer3.6-35B-A3B-UD-Q4_K_XL.ggufFor large-memory systems where preserving the merged model matters more than storage and transfer size.
QwiVer3.6-35B-A3B-UD-Q8_K_XL.ggufQwiVer is at its best when it is allowed to operate as more than a chat model. Long context, thinking preservation, native tool use and a small active footprint make it a natural fit for local agent and coding stacks.
Codebase navigation, multi-file reasoning, implementation, debugging and iterative changes where the model needs to stay coherent across a real project.
Tool-driven workflows, local shell agents, OpenCode-style coding systems, MCP environments and assistants that need to keep moving after the first answer.
262K native context gives the model room for large repositories, technical documents, logs and persistent reasoning state without immediately throwing information away.
Pair a QwiVer GGUF with the supplied BF16 multimodal projector for compatible image input while keeping the same BlackRiver language-model release.
The release chain is explicit. Qwen supplied the architecture and pretrained/post-trained foundation. BlackRiver built the curriculum adapter, selected Phase 10.2, merged it into BF16 and produced the MTP-preserving GGUF family.
Every public GGUF descends from the merged BlackRiver checkpoint.
Download the GGUF that fits your machine, keep the weights local and put a serious coding-and-agent model behind your own runtime. QwiVer is the BlackRiver answer to the idea that useful frontier-style behavior has to live behind somebody else's API.