deepseek-flash vs hy4-preview
Parameters, independent leaderboard coverage, real-task reports and open-weight availability are shown side by side; missing coverage is not a zero.
Parameters
| Comparison item | deepseek-flash | hy4-preview |
|---|---|---|
| Model Type | LLM | LLM |
| Context window | 1,000,000 tokens | — |
| Maximum output | 384,000 tokens | — |
| Input/output modalities | text, image → text | — |
| Endpoint | openai · openai-response · anthropic | anthropic · openai |
| Open weights | Available | Available |
| License | MIT | Apache-2.0 |
Leaderboard coverage
Every independent board covering either model is retained; an em dash means that board has not published a figure for that model.
Real-task reports
deepseek-flash
- RemakeBench Launch 005: four Flash models on 14 fixtures — Fourteen fixtures spanning game building, 3D scenes, CAD, drawing and robot control, run on Qwen 3.8 Flash, DeepSeek V4.1 Flash, GLM 5.3 Flash and Gemini 3.8 Flash. Finding: No model won overall. DeepSeek V4.1 Flash delivered the best submitted Turbofan CAD result (three valid native solids, with a wrong-sign coupling and LP/HP interference) and did not pass the single-turn Rubik's Cube task. Limitation: Pass/fail labels are video judgments without a machine-readable validator, and single attempts do not establish reliability; the lab publishes no aggregate score or overall winner. Several DeepSeek clips reuse earlier records that keep V4 labels and the Vision exp setting; the two results quoted here are campaign runs added for this comparison.
- MiMo-V2.6-Flash, DeepSeek V4.1 Flash and Grok 4.7 on one game prompt — Build a playable browser game from a single prompt with Command Code's /design command. Finding: DeepSeek V4.1 Flash scored 9/10 for $0.0089 with a playable one-shot game. MiMo-V2.6-Flash also scored 9/10 at $0.005; Grok 4.7 scored 8/10 at $0.20 and needed iterations. Limitation: One prompt rated by the author, whose profile lists work for Command Code, the tool that ran the test; there is no rubric and no repeated run.
- DeepSeek V4.1 Flash and GLM 5.3 Flash recreating two game menus — Recreate the Psychonauts 2 and Metroid Prime main menus as closely as possible to the originals. Finding: DeepSeek V4.1 Flash used 412k and 466k tokens, three to four times GLM 5.3 Flash's 134k and 111k, and the author judged GLM's menus clearly better on both games. Limitation: Which menu is closer to the original is the author's own judgement on two prompts, without a scoring rubric.
- Switching a coding project from GPT-6 Astra to DeepSeek V4.1 Flash — Review an existing project, list its issues and fix them. Finding: DeepSeek V4.1 Flash found several issues, asked which to tackle first and began fixing them; Kimi's review still found problems, which were passed back to DeepSeek to correct. Limitation: One uncontrolled self-report. The author ran DeepSeek through WorkBuddy's free access, so the remark about spending nothing extra does not reflect API pricing.
- Step 5 Preview and DeepSeek V4.1 Flash on a 3D arcade game — Build Katamari Tiny, a 3D arcade game, from a single prompt. Finding: DeepSeek V4.1 Flash used 26,898 tokens against Step 5 Preview's 39,830 and, in the author's view, was faster and produced the better game. Limitation: One prompt judged by an author who names DeepSeek V4.1 Flash as their default model; the post gives no timings or quality rubric.
- DeepSeek V4.1 Flash and GLM-5.3-Flash on a Blender geometry task, run locally — Write a Blender script that builds a dodecahedron trapped inside a dodecahedron. Finding: Both models produced a correct mesh with zero intersections. DeepSeek V4.1 Flash took 20 min 15 s and 62,928 tokens at 51.8 tok/s, about a fifth of GLM-5.3-Flash's time and half its tokens, and its computed clearance matched the measured mesh (0.238); the author found GLM's picture nicer. Limitation: One local run per model on different hardware and precision: GLM ran as a third-party EXL3 quantisation on two DGX Sparks and DeepSeek as FP8 on four, so speed and token counts describe those builds, not the official weights or a hosted API.
- Three Flash models build the same koi pond locally — A koi pond build, run as a local head-to-head. Finding: All three finished cleanly with zero fixes. DeepSeek V4.1 Flash was fastest at 21 minutes and 53.3K tokens, a third of GLM-5.3-Flash's time, and the author ranked it first. Limitation: One local run per model; the post states neither the prompt nor the quantisation or runtime used.
- Step 5 Preview and DeepSeek V4.1 Flash on one Three.js scene — A Three.js scene of giant red mesas in a desert, with long shadows and a huge empty sky. Finding: DeepSeek V4.1 Flash's scene was more restrained but had clearly better ground and mesa textures and a smoother camera; Step 5 Preview added more atmosphere, a tracked sun, particle and dust effects and on-screen text. Limitation: One prompt judged by eye by the author; the post does not say how either model was accessed.
- Hy4 Preview and DeepSeek V4.1 Flash on one voxel pagoda prompt — A small voxel pagoda garden scene in Three.js, in one HTML file. Finding: DeepSeek V4.1 Flash produced a much simpler scene than Hy4 Preview, whose shadow detail made its result look more like a finished photo. Limitation: One run per model, labelled by the author as personal experience rather than a benchmark, inside Tencent's own WorkBuddy product.
- 29 coding models on the same 60 tasks at Bito — The same 60 coding tasks at the author's company. Finding: DeepSeek V4.1 Flash scored 50.5/60 for $2.67, 93% of the top score (Claude Opus 5 at 54.5 for about $135) for 2% of the cost; a split sending only judgment-heavy tasks to Opus scored 53.5 for $18. Limitation: The author works at Bito, whose numbers these are; the title says over 1,500 tasks while the body describes the same 60 tasks, and the post does not describe the tasks or the scoring.
- AI Coding Daily LLM Coding Leaderboard — Implement and harden production-shaped projects in several stacks, including a Laravel app, a Go service and a React/TypeScript app, from the same prompts. Finding: DeepSeek V4.1 Flash scored 48.62 at max and 48.52 at high reasoning for $0.03–$0.05 per run in OpenCode, 23rd and 24th of 47 rows. Limitation: One author's benchmark whose code-quality half is graded by a judge model; rows run in different harnesses, the page is updated as models are added (read on 2026-09-24, 47 rows), and the author sells premium tutorials and advertising.
- kzhu's six-question LLM test leaderboard — Six questions scored out of 10: constrained sentence writing, two reasoning puzzles (replaced by pelican and compound-bow SVG animations in version 2.2), a browser OS with a space shooter, a 3D racing game and a Trello-style board. Finding: The released DeepSeek V4.1 Flash scored 49/60 on version 2.1 in Pi and 45 after re-running the two new version-2.2 SVG questions, where the pelican did not follow the ring and the compound bow had three structural errors (6 and 4); a pre-release build had scored 50. Limitation: One author's scores on six questions; rows use different harnesses and test versions (2.1 or 2.2), and a few rows' table total differs by one point from the author's note.
hy4-preview
- Hy4 Preview and DeepSeek V4.1 Flash on one voxel pagoda prompt — A small voxel pagoda garden scene in Three.js, in one HTML file. Finding: Hy4 Preview handled shadow detail and cast shadows clearly better, so its scene looked more like a finished photo; DeepSeek V4.1 Flash produced a much simpler scene. Limitation: One run per model, labelled by the author as personal experience rather than a benchmark, inside Tencent's own WorkBuddy product.
- kzhu's six-question LLM test leaderboard — Six questions scored out of 10: constrained sentence writing, two reasoning puzzles (replaced by pelican and compound-bow SVG animations in version 2.2), a browser OS with a space shooter, a 3D racing game and a Trello-style board. Finding: Hy4 Preview scored 48/60 on version 2.1 in WorkBuddy and 43 in Claude Code. In WorkBuddy it solved the lock puzzle and did well on the three coding tasks, but only half the constrained sentences passed and one 24-point answer broke the rules; in Claude Code it was very slow, repeatedly hit rate limits and left the racing game unfinished. Limitation: One author's scores on six questions; rows use different harnesses and test versions (2.1 or 2.2), and a few rows' table total differs by one point from the author's note.