DeepSeek V4 Flash vs Qwen3.8-Flash-Next vs GLM-5.3-Flash: the real leader in quality-to-price in 2026

In mid-to-late 2026 the “Flash” tier of open-weight models has become the practical workhorse for production systems: these are not the absolute frontier flagships (those still carry higher latency and much higher cost). They are the models teams actually route the majority of traffic to: high volume, strong enough capability, and dramatically better economics.

This article compares the three current leaders of that tier — DeepSeek V4 Flash (0731), Qwen3.8-Flash-Next (served in production as Qwen3.8-Flash), and GLM-5.3-Flash — using official model cards, Artificial Analysis independent measurements, provider pricing pages, and third-party reporting as of early September 2026. Every benchmark number below is sourced and dated; where a figure is vendor-reported rather than independently reproduced, we say so.

Model snapshot

Model	Developer	Architecture	Active Params	Context	Release	License
DeepSeek V4 Flash	DeepSeek	284B MoE	~13B	1M (up to ~1.31M on some providers)	Apr 2026; 0731 refresh Jul 31, 2026	Open weights (MIT)
Qwen3.8-Flash-Next	Alibaba (Qwen)	125B MoE + 51B n-gram embedding layer (~177B total)	~6B	262K native → 1M (YaRN)	Aug 26, 2026	Open weights (Qwen Community)
GLM-5.3-Flash	Z.ai (Zhipu)	320B MoE, hybrid linear + sparse attention	~18B	1,048,576 tokens	Aug 26, 2026	Open weights (MIT)

All three support tool use / function calling and reasoning modes (thinking is on by default in Qwen3.8-Flash-Next and DeepSeek V4 Flash). Qwen3.8-Flash-Next and GLM-5.3-Flash are natively multimodal — GLM accepts image and video input; DeepSeek V4 Flash is text-first in its main checkpoint (vision variants exist separately).

Published Benchmark Scores

This is the table the original version of this article was missing. Scores are as published by each vendor’s model card unless marked [AA], which denotes Artificial Analysis’s independent measurement. Harnesses, temperatures, and context limits differ per vendor — treat cross-model gaps of a few points as directional, not definitive.

How to read this table: GLM-5.3-Flash posts the strongest agentic-coding number of the tier (DeepSWE 63.4, vs GLM-5.2’s 46.2) and the top independent intelligence score. Qwen3.8-Flash-Next sweeps the language, reasoning, and office-agent rows — its CoWorkBench 73.9 against DeepSeek’s 45.1 is the single largest gap in the table — and does it with only 6B active parameters.

DeepSeek V4 Flash leads on nothing in this table except the security/repo niches, but it is competitive almost everywhere and, as the next sections show, wins decisively on speed and cost.

Benchmark	DeepSeek V4 Flash 0731	Qwen3.8-Flash-Next	GLM-5.3-Flash
AA Intelligence Index (independent)	50	56	57
Terminal-Bench 2.1 (terminal agent)	82.7 vendor / 79 [AA]	n.r.	84.3
DeepSWE 1.1 (agentic coding)	54.4	58.7	63.4
SWE-bench Pro (multilingual SWE)	56.0	62.5	n.r.
SWE-bench Multilingual	n.r.	81.0	n.r.
SWE-bench Verified	79.0 (April card, Max)	n.r.	n.r.
LiveCodeBench v6	90.6	91.9	n.r.
GPQA Diamond (reasoning)	90.8	91.7	n.r.
IFBench (instruction following)	79.2	81.3	n.r.
CoWorkBench (long-horizon office)	45.1	73.9	n.r.
JobBench (professional tasks)	41.3	55.7	n.r.
Toolathlon Verified (tool use)	70.3	73.5	n.r.
Cybergym (security agent)	76.7	n.r.	n.r.
NL2Repo (repo generation)	54.2	n.r.	n.r.
AutomationBench	25.1 (public)	n.r.	48.8
OfficeQA Pro	n.r.	n.r.	62.4
Agents’ Last Exam (pass@1)	25.2	24.3 (score 51.2)	n.r.
HLE (frontier reasoning)	33.8	35.9	55.3*
AndroidWorld (mobile use)	—	84.5	n.r.
MathVision (visual math)	—	95.7 (with code interp.)	n.r.
Z.ai Code Bench v1.0 (max)	n.r.	n.r.	29.0 (Opus 4.8: 29.5)

n.r. = not reported. *GLM’s HLE figure comes from a different, tool-augmented setup and is not comparable to the text-only runs of the other two.

Independent verification: Artificial Analysis

Vendor tables are marketing documents, so the independent column matters most. Artificial Analysis has measured all three models on its standardized harness:

Metric (AA, first-party API)	DeepSeek V4 Flash 0731 (max)	Qwen3.8-Flash-Next	GLM-5.3-Flash
Intelligence Index	50	56	57
Output speed	107.9 tok/s	87.8 tok/s	48.7 tok/s
Time to first token	1.50s	~competitive	1.52s
Tokens to complete the Index	~206M (down from 234M in April)	n.r.	~150M (class median: 110M)

Reference points: the open-weight class median output speed is ~67 tok/s, and GLM-5.3-Flash’s 57 on the Intelligence Index lands just 3 points behind the full GLM-5.3 (60) at roughly one-tenth the token price. DeepSeek’s 0731 checkpoint gained +10 points on the Index over the April preview purely from re-done post-training — same 284B/13B architecture. DeepSeek also holds the #2 open-weight spot on AA’s GDPval v2 Elo (1559), behind Kimi K3 (1687).

Note the verbosity column: GLM-5.3-Flash burns ~150M tokens to finish the Index (36% above the class median) and generates them at half the median speed. That is the “cheap tokens, expensive minutes” effect — decisive for interactive use, irrelevant for overnight batch jobs.

Pricing Snapshot (per 1M tokens, September 2026)

Model	Input	Cached input	Output	Blended (3:1 in:out)	Notes
DeepSeek V4 Flash	$0.14	$0.0028 (~98% off)	$0.28	~$0.18	OpenRouter providers from ~$0.03/$0.16; 2,500 concurrency limit (5× V4 Pro)
Qwen3.8-Flash-Next	$0.15	varies by provider	$0.47	~$0.23	Blended ~$0.23 measured; thinking tokens bill as output
GLM-5.3-Flash	$0.15	$0.03	$0.50	~$0.24	50% launch promo until Sep 9: $0.075/$0.25, cached $0.015 → blended ~$0.12

Cost per completed task tells the real story: on AA’s Intelligence Index, GLM-5.3-Flash runs about $0.09 per task at list price (~$0.045 on the promo tier) versus $0.68 for GLM-5.3 — and DeepSeek’s cost per task measured ~60% below GPT-5.6 Luna (max). One billing trap applies to all three: reasoning tokens are charged at output rates, so agentic workloads cost more than the headline input price suggests.

For detailed token math and tier comparisons, consult our model pricing benchmarks directory and see the GLM token pricing breakdown and Core plan savings to calculate operational break-even points under European Zero Data Retention.

Speed and Latency

The word “Flash” is doing different work in each name:

• DeepSeek V4 Flash — 107.9 tok/s on the first-party API, ~123 tok/s P50 across OpenRouter providers, 1.50s TTFT. The fastest of the tier by a wide margin, helped by DSpark speculative decoding.
• Qwen3.8-Flash-Next — 87.8 tok/s, comfortably above the ~67 tok/s class median. The best speed-per-intelligence point in the group.
• GLM-5.3-Flash — 48.7 tok/s despite a snappy 1.52s TTFT. Quick to start, slow to finish, and verbose on top. Fine for async pipelines; painful in interactive coding loops.

Why These Models Are So Efficient (Architecture Notes)

• Qwen3.8-Flash-Next is a public preview of the Qwen4 architecture: a 125B MoE paired with a 51B n-gram embedding layer (an idea borrowed from DeepSeek’s Engram paper) that acts as a cheap lookup memory, letting the model activate only ~6B parameters per token while matching 49B-active models on coding.
• GLM-5.3-Flash is the first GLM-5 model with hybrid attention (KDA linear layers interleaved with NoPE sparse MLA) plus IndexPool key compression: ~3× less attention compute and a 4.4× smaller KV cache than GLM-5.3 at million-token context. Trained on a 30T-token multimodal corpus; native FP8 weights (~306 GiB — self-hosting realistically requires an 8× Hopper-class node).
• DeepSeek V4 Flash 0731 changed no architecture at all — the entire +10 Intelligence Index jump came from re-running post-training (DeepSWE went 7.3 → 54.4 on identical weights). A reminder that in 2026, post-training is a release cycle.

Operational guidance by Use Case

Coding & code review — GLM leads agentic coding (DeepSWE 63.4, Terminal-Bench 84.3, within half a point of Opus 4.8 on Z.ai’s internal code bench). Qwen leads repository-level fixes (SWE-bench Pro 62.5, SWE-bench Multilingual 81.0). DeepSeek remains excellent value (SWE-bench Verified 79, Terminal-Bench 79 [AA]) at the lowest cost per token. Winner, agentic sessions: GLM. Winner, repo-scale fixes: Qwen. Winner, volume: DeepSeek.

Long-document analysis — All three reach ~1M tokens. GLM’s IndexPool was designed exactly for this; DeepSeek is proven at the extreme end; Qwen needs YaRN beyond 262K. Winner: GLM / DeepSeek tie.

Batch extraction & structured output — DeepSeek’s 108 tok/s + $0.0028 cached input is the default for high-volume pipelines (for GDPR compliance without China routing, deploy via the DeepSeek European sovereign API). Qwen (IFBench 81.3) and GLM are more accurate on complex schemas. Winner, volume: DeepSeek. Winner, accuracy: Qwen.

Multimodal / OCR / UI work — Only Qwen and GLM qualify natively. Qwen posts AndroidWorld 84.5 and MathVision 95.7; GLM takes image + video and powers Browser/Computer Use in ZCode, but trails Gemini 3.7 Flash on video benchmarks. Winner: Qwen for static visual tasks, GLM for video and agentic UI.

Interactive products (chat, pair-programming) — Throughput is the product here. DeepSeek or Qwen; GLM’s 49 tok/s will be felt by users. Winner: DeepSeek.

Overnight / async agentic jobs — GLM’s promo pricing (blended ~$0.12 until Sep 9) plus top agentic scores make it close to ideal. Winner: GLM.

Composite scores

Normalized 0–100, combining the benchmark table above, AA independent data, pricing, context, output capacity, recency, and versatility:

Metric	DeepSeek V4 Flash	Qwen3.8-Flash-Next	GLM-5.3-Flash
Capabilities	78	85	89
Pricing efficiency	94	82	88
Context window	95	88	95
Output capacity	92	80	85
Recency	75	95	95
Composite	84	86	90

• GLM-5.3-Flash (90) — strongest composite claim
• Qwen3.8-Flash-Next (86) — best efficiency story, most complete language/agentic table
• DeepSeek V4 Flash (84) — volume and speed king

A Necessary Caveat on Vendor Numbers

Every non-AA figure above is self-reported, and reproducibility gaps are already documented: AA’s independent Terminal-Bench 2.1 run scored DeepSeek 79% versus the vendor’s 82.7, and an independent DeepSWE re-test of DeepSeek’s V4 Pro measured pass@1 far below the official table. Qwen reports the highest score across two harnesses (Claude Code and mini-SWE-agent) for DeepSWE. GLM’s HLE 55.3 uses a tool-augmented setup. None of this means the numbers are fake — it means you should replicate the two or three benchmarks that map to your workload before committing.

---

---

Practical recommendation

• Best overall quality-to-price, especially async/agentic → GLM-5.3-Flash. Top independent intelligence (57), top agentic coding (DeepSWE 63.4), MIT weights, and promo pricing until Sep 9. Accept the 49 tok/s.
• Best speed + cost at volume → DeepSeek V4 Flash. 108+ tok/s, $0.14/$0.28 with near-free cache hits, 2,500 concurrency. The high-throughput default.
• Best balance and best architecture bet → Qwen3.8-Flash-Next. 6B active parameters delivering 56 on the AA Index, the strongest language/office-agent table, native multimodal, 88 tok/s.

The right engineering answer is routing, not loyalty: send simple high-volume traffic to the cheapest capable endpoint (usually DeepSeek), escalate agentic and multimodal work to GLM or Qwen, and re-run your own eval set every few weeks — this tier now moves on a monthly cadence.

---

FAQ

Which Flash model has the best quality-to-price in September 2026?
GLM-5.3-Flash for async and agentic workloads (AA Index 57 at ~$0.09/task list, ~$0.045 on promo). DeepSeek V4 Flash for sustained high-volume text. Qwen3.8-Flash-Next for the best all-round balance.

Is GLM-5.3-Flash actually fast?
No. Despite the name it generates ~49 output tokens/s — below the 67 tok/s class median and about half the flagship GLM-5.3’s speed — though its 1.52s time-to-first-token is competitive. Use it for batch, not interactive loops.

How long does the GLM-5.3-Flash promo last?
The 50% launch discount ($0.075 input / $0.25 output per 1M tokens) runs until September 9, 2026 on Z.ai’s API; some third-party gateways run parallel promos.

Can I self-host these models?
All three ship open weights. GLM-5.3-Flash needs ~306 GiB (FP8) and Hopper-or-newer GPUs for the current vLLM path. DeepSeek’s official recipe targets a 4× GB300 node. Qwen3.8-Flash-Next has an official FP8 variant, and community quantizations run on a single desktop (≈11 tok/s at 2-bit CPU, up to ~52 tok/s with GPU offload) — usable for evals, not production.

Which model is best for agentic coding?
On published numbers: GLM-5.3-Flash (DeepSWE 63.4, Terminal-Bench 84.3) for long agentic sessions; Qwen3.8-Flash-Next (SWE-bench Pro 62.5) for repository-level fixes; DeepSeek V4 Flash for cost-effective volume (SWE-bench Verified 79).

Should I trust vendor benchmark tables?
Directionally, yes; literally, no. Independent re-runs show gaps of ~4 points (Terminal-Bench) and occasionally more. Reproduce the benchmarks that match your workload on your own harness before locking a routing decision.

---