DeepSeek V4.1 Flash in CritiqueCode: A Completed Review Run
A 552B CED backbone, native vision, asymmetric 8B/16B activation, and a repaired CritiqueCode run that completed the review gate.
V4.1 Flash, tested
Authoring held. Review now completes.
critique.sh
DeepSeek V4.1 Flash is now an Inference option in CritiqueCode. The release-day author run made the correct one-line fix, stayed inside scope, and passed the fixture’s tests; direct checks also covered a tool call and native image input. The first /review exposed a controller contract bug, not a model verdict: nested review fields were underspecified. After tightening that contract, a standard-depth rerun completed all three specialist lanes with no failed cases and no promoted findings. That is a complete, cautious result — not a claim that one fixture proves correctness.
DeepSeek V4.1 Flash arrived on September 10, 2026 with a genuinely different compute shape. It is the first DeepSeek model built around the company’s Causal Encoder-Decoder architecture, and it is natively multimodal. The interesting question for a coding agent is not whether the launch table is large. It is whether that architecture survives a real tool loop, edits the right file, runs the checks, and returns machine-readable review output when the controller asks for it.
What 552B means — and what it leaves out
The clean description is a 552B-parameter language backbone, not a 552B total system. DeepSeek also describes 196B Engram parameters accessed through sparse token lookup. The backbone has 40 causal Transformer layers split into a 20-layer causal encoder and a 20-layer decoder. During prefill, 8B parameters activate per token; during decode, 16B activate. That asymmetric path is the core efficiency idea: spend less while ingesting a long repository context, then spend more while producing the answer or tool decision.
CED lets the decoder derive its global KV state from the encoder’s final hidden states. DeepSeek reports global runtime KV at 890 bytes per token — about one quarter of V4 Flash — while its separate persistent SSD or host-memory cache is about one eighth. Those are different cache layers and should not be collapsed into one slogan. For long agent sessions, both matter: HBM governs the live loop; persistent cache governs how cheaply the system can revisit old context.
deepseek/deepseek-v4.1-flash.The vendor benchmark story is strong, not universal
At maximum reasoning effort, DeepSeek reports 90.6 on Terminal-Bench 2.1, 74.2 resolved on DeepSWE v1.1, 54.8 on AutomationBench, 31.8 on Agent’s Last Exam, 78.9 on Chartography with tools, and 89.6 on BabyVision with tools. These are useful launch signals, but they are vendor-reported and max-effort results. Our run below tests the narrower question that matters in CritiqueCode: can the model complete an edit and return review data that the controller can actually validate?
Our direct inference smoke matrix
We called the release-day inference route from this repository with the local credential. Four functional paths completed: exact short text, TypeScript generation, a required function call, and native image input. The generated bracket parser compiled and passed five behavioral cases, including quotes and backslash escapes. The tool test selected get_weather with { city: "Dublin", unit: "celsius" }. The image test correctly identified a one-pixel white PNG.
The strict JSON-schema request was inconclusive because DeepInfra returned an upstream 429 on all three attempts. The required tool call also needed one retry after a 429. That is launch-day capacity evidence, not a model-quality failure, but it matters to agent operators because retries add wall time. Another budget probe used max_tokens: 32; all 32 tokens went to reasoning and the API returned no visible answer. Small output caps are unsafe when reasoning is enabled.
The CritiqueCode run: correct edit, then a completed review
The throwaway git fixture contained sumInvoice, one obvious quantity bug, and two tests. V4.1 Flash inspected three files, submitted a plan, changed one reduce expression, and ran both the focused and project test capabilities with exit code 0. The first author turn gave us the edit we expected. The important follow-up was making the review controller’s contract explicit instead of asking the model to infer nested shapes from a vague schema hint.
The complete model-authored change
One file, one write, both fixture tests passing.
- return lines.reduce((total, line) => total + line.price, 0)
+ return lines.reduce((total, line) => total + line.price * line.quantity, 0)The release-day /review failed at that schema boundary: all three specialist lanes returned shapes that missed required nested objects, even after a formatting retry. We tightened the lane prompts and adjudicator proof-plan instructions, then reran at standard depth. Correctness, contracts, and test-sufficiency all completed; two hypotheses reached adjudication, no case failed, and no finding was promoted. The recorded outcome is complete / none_promoted. That means the gate worked — malformed output stayed out of the result — and the controller fix removed the failure mode we could reproduce.
complete / none_promoted as a clean execution outcome rather than proof that every code path is correct.Run V4.1 Flash in CritiqueCode
CritiqueCode now presents DeepSeek V4.1 Flash as a selectable Critique Inference model. V4 Flash 0731 remains in the picker. The new entry declares text and image input, a 1,048,576-token context window, and a 384,000-token maximum output so the local Pi runtime does not artificially truncate the route.
Start a local author session
critique-code login
critique-code chat \
--models critique/deepseek/deepseek-v4.1-flash \
--cwd /path/to/repositoryOur recommendation is simple: use it for scoped authoring, native vision, and long-context agent loops; keep output headroom generous; and keep the review gate enabled. The repaired run shows that the model can participate in a structured review once the controller contract is explicit, while the no-finding outcome still deserves independent verification on your repository.