Y Recipes is the open systems library behind Y Computer: pinned models, optimized runtimes, exact-device benchmarks, raw evidence and reproducible builds for people who want to own the intelligence their work depends on.
We ran all 284.3 billion parameters of DeepSeek V4 Flash 0731 on one NVIDIA DGX Spark, then built a smaller Y speculative-decoding sidecar on that same machine.
| Exact same-machine result | Target only | Upstream DSpark | Y IQ3_M |
|---|---|---|---|
| Fixed 256-token generation | 16.93 tok/s | 28.09 tok/s | 28.29 tok/s |
| Mean wall time | 15.41 s | 9.40 s | 9.34 s |
| Sidecar size | — | 10.15 GiB | 7.95 GiB |
| Minimum observed memory available | 16.02 GiB | 7.01 GiB | 9.07 GiB |
| Integration gate | 6/6 | 6/6 | 6/6 |
That is 1.67× target-only throughput, 39.40% less wall time, a 21.64% smaller sidecar than upstream, and 2.06 GiB more observed memory headroom than the upstream path. The server was configured for 32,768 tokens; the longest tested input was 18,072 tokens; maximum process swap was 0 KiB.
- Inspect every result, request, hash and memory sample
- Build the Y IQ3_M sidecar
- Open the machine-readable summary
- See the visual benchmark on y.computer
The official DeepSeek checkpoint launches in the frontier agentic band. In DeepSeek's own same-panel evaluation, V4 Flash 0731 scores 82.7 on Terminal-Bench 2.1 versus 81.0 for GLM-5.2 and 85.0 for Claude Opus 4.8. It also lands between those two on DeepSWE, Toolathlon-Verified and Agents' Last Exam. See the official model card and full table.
That locates the official checkpoint, not this exact 3-bit target and Y sidecar. Speed is measured here; intelligence must be measured on a common harness. The exact Y stack is therefore getting a paired three-profile suite:
- Target-only
UD-IQ3_XXS. - The same target with the upstream DSpark sidecar.
- The same target with the Y
IQ3_Msidecar.
The first public quality panel will use pinned MMLU-Pro, GPQA Diamond, LiveCodeBench v6 and BFCL v4 prompts, seeds and output budgets across all three profiles. We will publish the scores, parse failures and raw outputs together. The existing 6/6 gate proves the serving stack works; it is not presented as an intelligence score.
Run one large OpenAI-compatible endpoint for coding agents, research, automations and internal tools without sending every prompt to a model host.
Open the DeepSeek-on-Spark build →
Connect a long-context agent endpoint to independent video-generation lanes and let one private system reason, plan and render.
Open the dual-Spark video factory →
Y tracks the newest downloadable weights, their real memory class, license, runtime and smallest useful device—from phone-class models to 284B desk-side systems.
The repository gives you the build. Y OS turns the build into a product:
- installs and pins the model, runtime and tools;
- exposes one private OpenAI-compatible endpoint;
- routes coding, research, media and automation apps to local models;
- manages updates, health, memory and recovery;
- keeps the full evidence record attached to the running configuration.
You can fork everything here and build it yourself. Or Y can ship the machine configured, benchmarked and ready to work.
Every published result names the model revision, runtime commit, exact hardware, prompt class, context, concurrency, run count and known boundary. Finished proof files include raw machine-readable outputs, hashes, commands, telemetry and a schema-valid manifest.
This repository distributes Y-authored documentation, harnesses and orchestration guidance—not third-party model weights. Models, projects and trademarks belong to their respective owners. Preserve upstream notices and review the applicable license before commercial deployment.