support for ds4 flash vision exp is available and https://github.com/antirez/ds4/blob/main/docs/MODELS.md talks about what model to download. we should check that it runs with our app, too, and is available. It needs a different dspark drafter https://github.com/antirez/ds4/blob/main/docs/SPECULATIVE_DECODING.md to support speculative decoding. The antirez/ds4 reference should have all the needed code already for it, so we should be able to just use it and it might be reduced to just having one more model in the catalog.
support for ds4 flash vision exp is available and https://github.com/antirez/ds4/blob/main/docs/MODELS.md talks about what model to download. we should check that it runs with our app, too, and is available. It needs a different dspark drafter https://github.com/antirez/ds4/blob/main/docs/SPECULATIVE_DECODING.md to support speculative decoding. The antirez/ds4 reference should have all the needed code already for it, so we should be able to just use it and it might be reduced to just having one more model in the catalog.
DeepSeek V4 Flash Vision Experimental is now available as deepseek-v4-flash-vision-exp, with its required vision encoder and separate Vision-Exp DSpark support artifact. All three artifacts have pinned sizes/hashes and strict checkpoint/tensor validation; the 0731 drafter cannot be paired with this model. The model is integrated with the model manager, settings, reasoning modes, API/evaluator, metrics, and session database.
The implementation includes Rust image preprocessing, the complete Rust-dispatched ViT/aligner graph, position-dependent image tokens, visual attention and MoE routing, user/tool images, multi-image conversations, persistent or transient encoder loading, resident and SSD execution, and warm/disk cache identity validation. No C or Python model/application runtime is called by the product.
Reviewed against DS4 0aaea5a238fb41a35106a551e73c8409dfb751ac (full reference SHA recorded in the evidence document). The review covered preprocessing/layout, encoder operations and BF16 rounding, visual attention/routing, continuation positions, cache invalidation, and checkpoint-specific DSpark. MTPLX is the Qwen oracle and has no corresponding DeepSeek functionality. The image-identity payload in DS4Server disk checkpoints is a DS4Server storage extension.
Validation completed in isolation:
Real SHA256 and metadata intake for all three GGUF files.
Real-weight encoder, numerical comparison with the published DeepSeek encoder, image-layout/attention vectors, routing and Metal attention primitives, wrong-checkpoint rejection, and tokenizer reasoning/continuation tests.
Full production generation/cache test in three fresh processes: resident (15.10 s), persistent encoder (15.94 s), and SSD with 4096-expert cache (31.46 s). Each passed red/blue recognition, ordered multiple images, user/tool image continuation, same-layout image replacement, warm/disk cache restoration, and text-only use afterward.
DSpark matched target-only output within every tested runtime mode, with real drafting: resident/persistent 15 drafted and 10 accepted; SSD 30 drafted and 0 accepted. Resident and SSD wording differs; no cross-mode token-equivalence claim is made.
Existing DeepSeek 0731 and GLM model/tokenizer regression controls passed.
Required gates passed: cargo fmt --all -- --check, cargo clippy --all-targets --all-features -- -D warnings, make bundle, and cargo test --all-features (323 application tests plus eight tool/supervisor tests; relevant ignored Metal/artifact tests run explicitly).
The SSD receipt retains existing GPU-copy preload mapped-range diagnostics; the existing host-load fallback succeeds. The encoder comparison is not a bitwise DS4 oracle claim. Detailed scope, reference pins, numerical results, reproduction instructions, and test receipts: docs/issue100-vision.
Implemented and pushed in commit [9bbe12e6b3a363e767d06d2417457af9740f9c63](https://git.rfc1437.de/hugo/DS4Server/commit/9bbe12e6b3a363e767d06d2417457af9740f9c63).
DeepSeek V4 Flash Vision Experimental is now available as `deepseek-v4-flash-vision-exp`, with its required vision encoder and separate Vision-Exp DSpark support artifact. All three artifacts have pinned sizes/hashes and strict checkpoint/tensor validation; the 0731 drafter cannot be paired with this model. The model is integrated with the model manager, settings, reasoning modes, API/evaluator, metrics, and session database.
The implementation includes Rust image preprocessing, the complete Rust-dispatched ViT/aligner graph, position-dependent image tokens, visual attention and MoE routing, user/tool images, multi-image conversations, persistent or transient encoder loading, resident and SSD execution, and warm/disk cache identity validation. No C or Python model/application runtime is called by the product.
Reviewed against DS4 `0aaea5a238fb41a35106a551e73c8409dfb751ac` (full reference SHA recorded in the evidence document). The review covered preprocessing/layout, encoder operations and BF16 rounding, visual attention/routing, continuation positions, cache invalidation, and checkpoint-specific DSpark. MTPLX is the Qwen oracle and has no corresponding DeepSeek functionality. The image-identity payload in DS4Server disk checkpoints is a DS4Server storage extension.
Validation completed in isolation:
- Real SHA256 and metadata intake for all three GGUF files.
- Real-weight encoder, numerical comparison with the published DeepSeek encoder, image-layout/attention vectors, routing and Metal attention primitives, wrong-checkpoint rejection, and tokenizer reasoning/continuation tests.
- Full production generation/cache test in three fresh processes: resident (15.10 s), persistent encoder (15.94 s), and SSD with 4096-expert cache (31.46 s). Each passed red/blue recognition, ordered multiple images, user/tool image continuation, same-layout image replacement, warm/disk cache restoration, and text-only use afterward.
- DSpark matched target-only output within every tested runtime mode, with real drafting: resident/persistent 15 drafted and 10 accepted; SSD 30 drafted and 0 accepted. Resident and SSD wording differs; no cross-mode token-equivalence claim is made.
- Existing DeepSeek 0731 and GLM model/tokenizer regression controls passed.
- Required gates passed: `cargo fmt --all -- --check`, `cargo clippy --all-targets --all-features -- -D warnings`, `make bundle`, and `cargo test --all-features` (323 application tests plus eight tool/supervisor tests; relevant ignored Metal/artifact tests run explicitly).
The SSD receipt retains existing GPU-copy preload mapped-range diagnostics; the existing host-load fallback succeeds. The encoder comparison is not a bitwise DS4 oracle claim. Detailed scope, reference pins, numerical results, reproduction instructions, and test receipts: [docs/issue100-vision](https://git.rfc1437.de/hugo/DS4Server/src/commit/9bbe12e6b3a363e767d06d2417457af9740f9c63/docs/issue100-vision).
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
support for ds4 flash vision exp is available and https://github.com/antirez/ds4/blob/main/docs/MODELS.md talks about what model to download. we should check that it runs with our app, too, and is available. It needs a different dspark drafter https://github.com/antirez/ds4/blob/main/docs/SPECULATIVE_DECODING.md to support speculative decoding. The antirez/ds4 reference should have all the needed code already for it, so we should be able to just use it and it might be reduced to just having one more model in the catalog.
Implemented and pushed in commit 9bbe12e6b3a363e767d06d2417457af9740f9c63.
DeepSeek V4 Flash Vision Experimental is now available as
deepseek-v4-flash-vision-exp, with its required vision encoder and separate Vision-Exp DSpark support artifact. All three artifacts have pinned sizes/hashes and strict checkpoint/tensor validation; the 0731 drafter cannot be paired with this model. The model is integrated with the model manager, settings, reasoning modes, API/evaluator, metrics, and session database.The implementation includes Rust image preprocessing, the complete Rust-dispatched ViT/aligner graph, position-dependent image tokens, visual attention and MoE routing, user/tool images, multi-image conversations, persistent or transient encoder loading, resident and SSD execution, and warm/disk cache identity validation. No C or Python model/application runtime is called by the product.
Reviewed against DS4
0aaea5a238fb41a35106a551e73c8409dfb751ac(full reference SHA recorded in the evidence document). The review covered preprocessing/layout, encoder operations and BF16 rounding, visual attention/routing, continuation positions, cache invalidation, and checkpoint-specific DSpark. MTPLX is the Qwen oracle and has no corresponding DeepSeek functionality. The image-identity payload in DS4Server disk checkpoints is a DS4Server storage extension.Validation completed in isolation:
cargo fmt --all -- --check,cargo clippy --all-targets --all-features -- -D warnings,make bundle, andcargo test --all-features(323 application tests plus eight tool/supervisor tests; relevant ignored Metal/artifact tests run explicitly).The SSD receipt retains existing GPU-copy preload mapped-range diagnostics; the existing host-load fallback succeeds. The encoder comparison is not a bitwise DS4 oracle claim. Detailed scope, reference pins, numerical results, reproduction instructions, and test receipts: docs/issue100-vision.