Research via TinyFish, 2026-09-11
Side-by-side
| DeepSeek V4.1 Flash | GLM 5.3 Flash | |
|---|---|---|
| Released | Sep 10, 2026 | Aug 26, 2026 |
| Architecture | Causal Encoder-Decoder (CED), sparse MoE | Hybrid sparse + linear attention (KDA + NoPE Sparse MLA), MoE |
| Total params | 552B backbone (+196B Engram) | 320B |
| Active per token | 16B (8B prefill / 16B decode) | 18B |
| Layers | 40 (20 encoder + 20 decoder) | 45 (routes 8 of 288 experts per token) |
| Input price | .30/M peak, .15/M off-peak | .15/M list, .075/M promo |
| Output price | .20/M peak, .60/M off-peak | .50/M list, .25/M promo |
| Cache read | .003/M | .03/M list, .015/M promo |
| Context | 1M tokens (384K max output) | 1M declared (1.3M on some providers) |
| Throughput | ~190 tps (AA) | ~71 tps (AA), ~49-99 tps OR |
| TTFT | 1.00s (AA) | 1.46-2.07s |
| Cost per task (AA) | .27 | .045 discounted / .09 |
| Modalities | Text + image in, text out | Text + image + video in, text out |
| Providers (OpenRouter) | 4 | 25 |
| License | MIT, open weights | MIT, open weights |
| Weights on disk | 166.9 GB (V4 Flash, FP4+FP8) | ~306 GiB FP8 |
| Self-host floor | 2x H200 (V4 Flash) | 8-GPU Hopper-class node |
Composite Intelligence Scores (same index v4.3)
| Source | DeepSeek V4.1 Flash | GLM 5.3 Flash |
|---|---|---|
| Artificial Analysis Intelligence Index v4.3 | 40 | 42 |
| AA predecessor V4 Flash 0731 (v4.3) | 50 | - |
| AA predecessor V4 Flash 0731 (v4.1.1) | 52 | 57 |
| BenchLM category aggregates | Agentic 84.4 / Knowledge 56.6 / Multimodal 73.4 | 61.69/100, rank #59, 90% CI 50.2-73.2 |
| Z.ai in-house code bench | - | 29.0 (vs Opus 4.8 at 29.5) |
On the same v4.3 index the two models are essentially tied: 40 vs 42. The earlier 57-vs-52 gap was v4.1.1 only; on v4.3 GLM falls to 42 while V4.1 Flash lands at 40. Both are far above the open-weight median of 18.
Coding Benchmarks (verified rows)
| Benchmark | DS V4.1 Flash | GLM 5.3 Flash | Winner |
|---|---|---|---|
| Terminal-Bench 2.1 | 90.6 | 84.3 | DeepSeek |
| Terminal-Bench 3.0 | 30.0 | - | - |
| Terminal-Bench 4.0 | 31.2 | - | - |
| DeepSWE v1.1 | 74.2 | 63.4 | DeepSeek |
| NL2Repo | 65.4 | 56.3 | DeepSeek |
| ProgramBench | 20.3 | - | - |
| Codeforces rating | 3471 | - | - |
| LiveCodeBench (base) | 29.9 (V4 Flash) | 37.6 | GLM |
V4.1 Flash leads every shared coding row, by 6 points on TB2.1 and 11 on DeepSWE. TB 3.0/4.0 are far harder (frontier models score 43-52 on TB 4.0); V4.1 Flash still leads GLM-5.3-tier models there.
Agentic Benchmarks
| Benchmark | DS V4.1 Flash | GLM 5.3 Flash | Winner |
|---|---|---|---|
| AutomationBench v1.0.6 | 54.8 | 48.8 | DeepSeek |
| Toolathlon Verified | 78.4 | 78.4 | Tie |
| Agents' Last Exam | 31.8 | 26.3 | DeepSeek |
| HLE w/ tools | 63.9 | 55.3 | DeepSeek |
| GDPval-AA v2 (Elo) | - | 1773 (leaderboard top) | GLM |
| CyberGym | 88.1 | 88.0 | ~Tie |
| ExploitGym | 15.3 | - | - |
| SEC-Bench Pro | 62.8 | - | - |
GLM's standout agentic result is GDPval-AA v2 Elo 1773 (top of that leaderboard) and the 48.8 vs 41.0 lead over Opus 4.8 on AutomationBench. DeepSeek wins the head-to-head on everything it has published for V4.1 Flash, though the two labs use different harnesses, so cross-vendor rows are directional.
Knowledge / Reasoning / Math
| Benchmark | DS V4.1 Flash | GLM 5.3 Flash |
|---|---|---|
| GPQA Diamond | 90.9 | - |
| HLE (text-only) | 39.1 (36.8 full) | - |
| MMLU (base) | 88.5 (V4 Flash) / 74.1 MMLU-Pro | 88.1 |
| SimpleQA (base) | 31.2 | 33.5 |
| BBH (base) | 84.9 | 86.6 |
| MathArena Apex | 65.6 | - |
| HMMT Feb 2026 (base) | 94.8 | - |
Multimodal / Vision
| Benchmark | DS V4.1 Flash | GLM 5.3 Flash |
|---|---|---|
| CharXiv Reasoning w/ tools | - | 89.4 (Opus 89.9) |
| Chartography w/ tools | 78.9 | 78.0 |
| BabyVision | 89.6 | 53.4 |
| ZeroBench w/ tools | 49.0 | - |
| MMVU | - | 80.5 |
| MVbench (video) | - | 77.8 |
| OSWorld 2.0 | - | 59.1 |
| Vision2Web | - | 77.8 |
| OfficeQA Pro | - | 62.4 |
| ExtractBench mean / short / medium | - | 80.75 / 96.30 / 51.56 |
GLM supports video input; V4.1 Flash is image+text only. On static-chart tasks they are neck-and-neck (V4.1 Flash slightly ahead on Chartography and BabyVision); GLM leads on GUI/OS/computer-use evals (OSWorld, Vision2Web) where it has data and DeepSeek does not.
Speed (Artificial Analysis first-party API)
| Metric | DS V4.1 Flash | GLM 5.3 Flash |
|---|---|---|
| Output speed | 190.1 tps | 71.4 tps |
| TTFT | 1.00s | 1.46-2.07s |
| Cost to run AA index | .89 | .28 |
| Verbosity on AA index | 250M tokens (median 130M) | 180M tokens (median 130M) |
V4.1 Flash is ~2.7x faster than GLM on hosted APIs. Both are verbose; GLM cannot disable thinking (only reduce via reasoning_effort), while V4.1 Flash offers effort 1-100.
Pricing detail
- GLM list: .15 in / .03 cached / .50 out per 1M. Promo until Sep 9, 2026: .075 in / .015 cached / .25 out.
- Gateways: EmpirioLabs .07/.25, Venice .09/.31, Requesty .14/.45 (55% off promo to Sep 9), Qubrid .0863 in.
- DeepSeek: peak vs off-peak (off-peak = 50% of peak). V4.1 Flash off-peak .075/.30, peak .15/.60.
- AA blended: ~.10/M GLM vs ~.18/M (V4.1 Flash).
Hardware
- DeepSeek V4 Flash (0731): 166.9 GB FP4+FP8 checkpoint, runs on 2x H200.
- DeepSeek V4.1 Flash: KV cache 890 B/token = ~4x smaller than V4 Flash; sizing still being published.
- GLM 5.3 Flash: ~306 GiB FP8, 8-GPU Hopper-class floor; flat 4-bit still ~160 GB.
Caveats
- No vendor table includes the other model; only third-party rulers (Artificial Analysis, BenchLM) enable like-for-like reads.
- Harness/temperature/versions differ per vendor, so decimal-level comparisons are directional.
- V4.1 Flash released Sep 10; AA index (v4.3) = 40, GLM = 42 on the same index.
- GLM attention drift reported beyond ~700K tokens.
- AA GLM score of 42 is on v4.3; the 57 figure was v4.1.1. Do not compare 40 (v4.3) against 57 (v4.1.1).
Bottom line
- On the same v4.3 Intelligence Index the two models are essentially tied (40 vs 42).
- On every shared coding/agentic benchmark where both have numbers, DeepSeek V4.1 Flash matches or beats GLM 5.3 Flash while being ~2.7x faster - but it is 2x the price at GLM promo rates and at list price is the same input cost with a higher output cost.
- GLM 5.3 Flash edge: video input, GDPval/tool-use GUI strengths (OSWorld, Vision2Web, Toolathlon parity), much cheaper per task (.045-0.09), and a settled independent score.
- Practical split: V4.1 Flash through the DeepSeek API for coding agents, long-horizon terminal work, and speed-critical streaming; GLM 5.3 Flash for high-volume multimodal extraction/browser-use agents and cheapest-per-task defaults.
Sources
- Artificial Analysis: deepseek-v4-1-flash, glm-5-3-flash model pages + comparison
- BenchLM: deepseek-v4-1-flash profile, v4-flash-0731-vs-glm-5-3-flash comparison
- Z.ai blog (official benchmarks), DeepSeek model card / tech report (HF)
- DataCamp, Qubrid, Ampere.sh, Yotta Labs, local-ai-zone analyses
- OpenRouter model pages (pricing/providers/throughput)