FULL STORY

DeepSeek V4.1 Flash: From Limited Beta to Open-Source Release

DeepSeek V4.1 Flash went from a limited beta leak on Sept 8 to an official open-source release on Sept 10, topping open-weight leaderboards with near-frontier performance at a fraction of the cost.

2026-09-08 ~ 2026-09-12 · 18 episodes · 211 posts

Episode 1 · DeepSeek V4.1 Flash Opens Limited Internal Beta with New Architecture and Native Multimodality (2026-09-08, 18 posts)

Starting September 8, multiple sources reported that DeepSeek quietly opened a limited internal beta for V4.1 Flash, a "middle version" with model ID deepseek-v4.1-flash-expires-on-0910, the slug indicating expiry on September 10. Key upgrades include a brand-new architecture and native multimodality, with the official line claiming stronger capability, faster speed, and lower cost. Calling conventions, billing, and concurrency rules are now largely clear, developer benchmarks have emerged, and the model has appeared on Hugging Face, though an official public announcement and full evaluations remain pending.

Confirmed

  • Jiqizhixin reported that DeepSeek officially announced the internal test of the V4.1 Flash middle version; APPSO, Zhidongxi and other outlets confirmed the same.
  • Usage: keep baseurl unchanged and set the model name to an ID beginning with deepseek-v4.1-flash-expires-on-09 (e.g., deepseek-v4.1-flash-expires-on-0910), where the date marks expiry.
  • According to vista8's relay of an official WeChat group notice, billing matches V4 Flash with max 20 concurrent requests per account; teortaxesTex's leaked info points to the same concurrency limit and a vision-capable build (V4-Flash-Vision [Intermediate]).
  • On pricing, MiaAIlab (relayed by MicahBerkley) described the price as extremely aggressive: roughly 58 million tokens per US dollar, available for about 2 days.
  • Third-party account LuminaBench spotted the intermediate build being tested via API; tokenbender suggested it precedes the official release.
  • On September 10, Kevin Kern tested the preview endpoint, measuring 300 tok/s and heavily parallelizing subagents; he noted the preview endpoint does not yet accept vision input. Zhidongxi reported output speeds up to 507 tokens/s.
  • On September 10, Reddit users found a deepseek-ai/DeepSeek-V4.1-Flash model page on Hugging Face, suggesting a quiet official upload; specs unconfirmed.

Unconfirmed

  • The news initially spread via bloggers such as op7418 and AGI Hunt from a DeepSeek WeChat group, unverified at the time; Jiqizhixin later framed it as officially announced, but a formal public announcement and full capability details are still pending.
  • @NFTChen (relayed by solyarisoftware) claimed V4.1 Flash would be officially released around September 10 Beijing time, with V4 Pro requests routed at Flash pricing; not confirmed by DeepSeek.
  • The preview endpoint's lack of vision input sits in tension with the "native multimodality" claim; whether the formal release adds vision remains to be seen.
  • teortaxesTex relayed community banter hoping it can at least beat GLM 5.3 Flash outside of vision—opinion only, no benchmark basis.

Why it matters

  • This is DeepSeek's first reported use of an "expires-on" middle-version limited beta with explicit billing and concurrency rules plus aggressive pricing, signaling a product cadence of rapid iteration through real developer workloads before formal release.
  • Native multimodality marks a major upgrade to DeepSeek's flagship line; with 300 tok/s measured, an appearance on Hugging Face, and the community benchmarking it against GLM 5.3 Flash, the formal release's multimodal capabilities deserve close attention.

Episode 2 · DeepSeek V4.1 Flash Tested: Blazing 350 Tokens/s but Still Experimental (2026-09-08, 2 posts)

Hands-on tests of DeepSeek's experimental V4.1-Flash model show decoding speeds of roughly 300-400 tokens per second at competitive prices, though reviewers note it remains early-stage and prone to overthinking.

Episode 3 · DeepSeek cuts V4-Flash API prices with new peak/off-peak billing from Sept 10 (2026-09-08, 6 posts)

DeepSeek officially announced a price cut for its V4-Flash (flash) API series effective Sept 10, 12:00 Beijing time, restoring input prices to pre-hike levels and switching to peak/off-peak billing, with the largest discounts in off-peak hours. The new scheme replaces the expiring "Intermediate" V4.1 pricing; the price sheet had leaked via community channels a day earlier.

Confirmed

  • DeepSeek platform announcement: V4-Flash price cut from Sept 10, 12:00 (reported by 机器之心, 量子位, aigclink).
  • Off-peak pricing (yuan per million tokens): cache-hit input 0.02 (down 60% from 0.05); cache-miss input 1 (down 33.33% from 1.5); output 4 (down 11.11% from 4.5).
  • Peak hours are weekdays 9:00–12:00 and 14:00–18:00; the rest is off-peak. Peak prices are higher, per the official schedule.
  • The new scheme replaces the expiring "Intermediate" V4.1 pricing (per teortaxesTex).
  • michaelsoftbinbows noted that DeepSeek V4 Flash (0731) via openinference on OpenRouter costs only $0.05 input / $0.16 output per 1M tokens off-peak, below official pricing (from $0.22), drawing wide discussion.

Unconfirmed

  • WeChat account 赛博禅心 mentioned a 20-concurrency limit accompanying the new scheme, flagged as pending official confirmation; the official announcement is silent on this.

Timeline

  • Sept 8: teortaxesTex summarized the leaked pricing; 赛博禅心 posted the cut with effect time and concurrency details, unconfirmed.
  • Sept 9: Official announcement confirmed via 量子位, 机器之心, aigclink; OpenRouter third-party low pricing sparked discussion.
  • Sept 10, 12:00: New prices take effect.

Why it matters

  • Input prices returning to pre-hike levels partially rolls back the earlier increase, directly benefiting developers and vendors relying on DeepSeek APIs.
  • Off-peak cache-hit pricing as low as 0.02 yuan per million tokens, combined with time-based billing, will push users toward off-peak calls and higher cache utilization, reflecting shifts in DeepSeek's capacity and scheduling.
  • The notable gap between official and third-party channels (e.g., OpenRouter) offers developers a cheaper alternative access path.

Episode 4 · DeepSeek V4.1 Flash Nears Frontier Models at a Fraction of the Cost, Independent Tests Show (2026-09-09, 15 posts)

DeepSeek V4.1 Flash, along with the V4.1 Flash Vision Beta opened the same day, has undergone multiple rounds of independent testing since release. The consensus is that it approaches—and in some cases matches—frontier closed models at roughly 1/30 to 1/70 of their cost, corroborated by Artificial Analysis ranking it 6th of 113 models.

Confirmed

  • Artificial Analysis benchmark (reasoning version, Max Effort): Intelligence Index 40, ranked 6th of 113 models, 190 tokens/s throughput; open weights, released September 2026.
  • Design comparison by NFTChen: 81.2 points, 98% of GPT-6 Astra's 82.7, above Claude Fable 5.1's 80.3, at 1.4% of Astra's cost; similar figures reported by OpenDesign Arena (via uxl) and OpenDesignHQ (via ChrisUniverse, authenticity unverified).
  • Security benchmark by pilvar222: re-discovers 65.6% of recent CVEs in a single run (prior version 55.2%), 84.4% at pass@3 (prior 75%), matching frontier models.
  • Bug-fixing test by Pawel Huryn: 105 hidden real bugs across 2 repos; V4.1 Flash reaches the Pareto frontier at 3.5% of Opus's cost (1/28). Reference: Opus 5 (max) 27 bugs at $51.33, Grok 4.6 (max) 27 at $16.96; sander's retelling cites 24 points for V4.1 Flash.
  • Knowledge-work evaluation by himanshustwts: 22 internal tasks across engineering, operations, finance, healthcare; 15/21 passes, near SOTA at 2% of competitors' cost; entire run cost $0.34 with GPT-5.5 as judge and 120+ validators per task.
  • Vision: Reddit user cheezeerd compared V4.1 Flash Vision Beta vs V4 Flash Vision on 5 tasks on launch day; new version clearly better and more reliable, with lower API pricing.
  • Terminal agents and pricing: via Cline (retold by dejavucoder), V4.1-Flash surpassed V4-Pro 0813 on Terminal-Bench within a month of release; pricing from $0.30 per million input tokens, 1/20 of Opus.
  • Local deployment: a developer (via TheZachMueller) ran full-precision DeepSeek 4.1 Flash with the DSpark engine at 200 TPS on 4 Apple Max-Q chips with only 64GB system memory, key trick being handling of the 200GB Engram.

Unconfirmed

  • Apart from the Artificial Analysis ranking, all results come from third-party individual or team tests with varying samples and methodologies; OpenDesignHQ's benchmark is unverified; the Terminal-Bench claim is second-hand via Cline; the Mac 200 TPS figure is a single developer's report without official metrics or large-scale replication.

Why it matters

  • Independent tests across diverse task types mutually corroborate the cost-performance claim rather than relying on a single leaderboard, with Artificial Analysis providing neutral quantitative backing.
  • If reproducible, a near-frontier model at ultra-low prices would compress frontier closed-model pricing and lower migration barriers for developers and enterprises.
  • Full-precision high-speed operation on consumer Apple hardware adds appeal for local, private deployment scenarios.

Episode 5 · DeepSeek Releases Open-Source V4.1 Flash, Beating Its Own Flagship at a Fraction of the Cost (2026-09-09, 57 posts)

DeepSeek officially released and open-sourced V4.1 Flash on September 10 across web, app, and API, with weights on Hugging Face under the MIT license (fp8/8-bit builds, transformers support) plus an accompanying paper. The company claims it beats its own flagship V4 Pro on performance, cost, and speed, making it one of the most-watched open-source releases of the season.

Confirmed

  • Architecture: A 552B-parameter asymmetric MoE with a new Causal Encoder-Decoder design—8B parameters active on input, 16B on output—and up to 1M-token context. Cross-layer shared compressed KV cache saves 7/8 of KV cache cost (@YuchenjUW notes a further 4x per-token KV reduction), plus a two-stage sparse indexer.
  • Multimodal & ecosystem: First model in the new architecture family with native vision (image-text-to-text), and the smallest member so far, paving the way for larger siblings. vLLM and SGLang offer day-zero support verified on NVIDIA/AMD GPUs; DeepSeek is soliciting 2000-GPU-plus-storage cluster-scale deployment partnerships.
  • Pricing & benchmarks: New pricing with 50% peak/off-peak gap; community reports $0.14-0.15/M input and $0.60/M output. @kimmonismus notes output price is 6% of GPT-5.6 Sol's; @YuchenjUW puts total cost near 3%. Per Artificial Analysis, it scores 40 on the Intelligence Index—beating DeepSeek's 1.6T Pro flagship—at roughly 4x lower token price. Community posts cite DeepSWE v1.1 at 74.2%, CyberGym 88.1 vs 84.5, and API cost around 2.5% of Opus 5.
  • V4 Pro retirement: Until V4.1 Pro launches, all V4 Pro requests are auto-routed to V4.1 Flash at lower prices; per @markk, routing starts September 14. V4-Flash and V4-Flash-Vision-E are also being retired. This comes only 6 weeks after July's V4-Flash.
  • Reception: Hugging Face co-founder Thom Wolf says it tops the open-source leaderboard at rock-bottom prices and shared a forward-pass visualization video; antirez is more skeptical, seeing only a marginal gain over 4.0 Flash despite doubled backbone parameters.

Unconfirmed

  • Some reposts (@solyarisoftware) cite 769B total parameters, contradicting the official 552B figure. @ns123abc's claims of 1/86 competitor cost and 420-507 tok/s output speed stem from unverified leaks.

Why it matters

  • Matching or beating a flagship at a fraction of the cost, fully MIT-licensed, could further compress industry inference prices and pressure closed vendors—though the generational jump over its predecessor, as antirez notes, remains debated.
  • V4 Pro users should watch the September 14 auto-routing cutover and evaluate real-world behavior beforehand.
  • 1M-token context, native vision, and 7/8 KV cache savings directly affect long-document and agentic workflows, and the model sets the stage for larger siblings in the same family.

37 more related posts →

Episode 6 · DeepSeek V4.1 Flash Leaks: 552B Asymmetric MoE Claimed to Rival GPT-5.6 Sol (2026-09-10, 21 posts)

Starting September 10, multiple posters leaked claimed official benchmarks and technical details of DeepSeek V4.1 and V4.1 Flash, with Sentdex reporting the next day that weights had been released. If true, V4.1 would challenge GPT-5.6 Sol on agentic/coding benchmarks with a 552B asymmetric-activation architecture, dramatically cut inference memory and cost via extreme KV-cache compression, and introduce a rare 1-100 continuous reasoning-effort control. Nothing is officially confirmed, and iScienceLuvr among others suspects benchmark gaming.

Confirmed

  • All information is unofficial leaks, from teortaxesTex, eliebakouch, ChrisGPT, nrehiew, iScienceLuvr, AdinaYakup, zainhas, Sentdex, @sheriyuo (via AndrewZhao) and @MiaAIlab (via airesearch12), with figures cross-corroborated; SemiAnalysis also posted "congratulations" (via bodonoghue85).
  • Total parameters 552B with a new Causal-Encoder-Decoder architecture: 8B input activation / 16B output activation; nrehiew notes the same model activates different parameter counts across prefill and decode stages, an unusual design. Sources stress costs well below same-size models; sheriyuo claims it is smarter than DeepSeek's own flagship; MiaAIlab says capability is on par with or better than GPT-5.6 Sol, Claude Opus 5, GLM 5.3 and Kimi K3; Sentdex's cited post says weights are out and it beats Opus-5, GPT-5.6-Sol, Kimi-K3 and GLM-5.3 in key areas.
  • Leaked benchmarks: TerminalBench 4.0 at 31.2, CyberGym 88.1 (vs GPT-5.6 Sol's 84.5), HLE 63.9; ChrisGPT claims leads across agentic/coding benchmarks; nrehiew reports V4.1 Flash tops DeepSWE at 74.2, above Opus 5 and Gemini 3.8 Flash.
  • Per eliebakouch, V4.1 Flash is a natively multimodal model trained on 45T tokens, with an architecture praised as the most novel in years; AdinaYakup notes native vision is merged into one endpoint.
  • zainhas highlights reasoningeffort continuously adjustable from 1 to 100, rather than discrete low/medium/high tiers, plus a rare API compatibility mapping — a design they say they have never seen before.

Architecture and engineering highlights

  • Per teortaxesTex, KV cache compressed to 890 bytes/token (under 1GB per million tokens); HBM demand down another 3.9x and SSD down 8x versus V4-Flash. AdinaYakup cites KV cache at 1/437 of the first generation, HBM at 1/4 and SSD at 1/8 of the previous generation.
  • Scobleizer's relay adds pricing: $0.15 input / $0.60 output per million tokens; AdinaYakup calls it one-fortieth of Opus's price.
  • Also per teortaxesTex (via zephyrz9), V4.1 removes V4's stopgaps: drops HCA, generalizes CSA into a more general primitive, removes sparse-attention warmup, and adopts a simpler single-pass mHC; a new sparse attention design is also mentioned (via donglixp).
  • teortaxesTex also jokingly relayed a report signed "fruit fly connectome" claiming V4.1-Flash inference optimized on 2× NVIDIA GB300 DGX Stations with 56k tok/s prefill and 201.7 tok/s c1 decode; its credibility is dubious by nature.

Unconfirmed

  • All benchmarks, specs, pricing, throughput figures, architecture details and the weights-release claim lack official sources; iScienceLuvr explicitly doubts the "GPT-5.6 Sol level" scores, suspecting benchmark gaming, with "benchmarkmaxxed" skepticism appearing in related threads.

Why it matters

  • If true, V4.1 goes head-to-head with GPT-5.6 Sol on agentic/coding while asymmetric activation and extreme KV compression slash deployment memory barriers; the 1-100 continuous reasoning effort would offer a new interaction paradigm for API-side inference cost and depth control, with significant cost implications in an open-weights ecosystem.

1 more related posts →

Episode 7 · DeepSeek V4.1 Flash Benchmark Results Leak, Sparking Community Discussion (2026-09-10, 2 posts)

Early benchmark screenshots circulating on Reddit show DeepSeek V4.1 Flash achieving SOTA on the deepswe leaderboard alongside sol and fable5, with competitive terminal bench results, sparking community discussion.

Episode 8 · DeepSeek Shrinks Per-Token KV Cache 54x in Nine Months to Sub-KB Levels (2026-09-10, 6 posts)

According to X users zephyrz9 and teortaxesTex, DeepSeek compressed its per-token KV cache by 54x over roughly nine months, reaching sub-KB levels of about 890 bytes per token, drawing wide community attention to its inference efficiency and memory optimization.

Confirmed

  • zephyrz9 reported the 54x per-token KV cache compression over nine months, reflecting rapid iteration in inference efficiency and memory optimization.
  • teortaxesTex cited the specific figure: about 890 bytes per token on the new model, an unusually low footprint.
  • kalomaze confirmed single-token KV storage is now under 1KB, consistent with DeepSeek's efficiency-focused track record, but noted the associated training and serving infrastructure is "agonizing" in cost and engineering complexity.
  • teortaxesTex argued that an architecture-level sparsification change is highly significant, boosting token efficiency toward OpenAI-style territory and away from massive parameter-activation waste.
  • Investor 0xdoug called DeepSeek definitively the third frontier lab after OpenAI and Anthropic, with a different path: OpenAI and Anthropic push the intelligence frontier at acceptable cost, while DeepSeek pushes down cost at acceptable intelligence.

Unconfirmed

  • teortaxesTex's claim that the new release carries "ChatGPT fingerprints" is personal interpretation and unverified.
  • Whether long-context capability holds under such extreme KV compression remains unproven; teortaxesTex explicitly awaits long-context benchmark results.

Why it matters

KV cache is the dominant source of memory usage in long-context inference; a 54x reduction means longer contexts and higher concurrency on the same hardware, directly cutting inference costs. The "third frontier lab" framing underscores DeepSeek's efficiency-first differentiation — but if it comes at the cost of long-context quality or extreme infrastructure complexity, the value proposition still needs benchmark data and deployment practice to verify.

Episode 9 · Leaked DeepSeek-V4.1-Flash Touted as New Open-Source King (2026-09-10, 3 posts)

DeepSeek-V4.1-Flash leaked on Hugging Face under MIT license, reportedly beating GLM 5.3 and Kimi K3 at 4-10x lower cost ($0.3/M input, $1.2/M output), ranking sixth on Codeforces.

Episode 10 · Bug Hunt Bench Ranks Frontier Coding Models on 105 Real Bugs (2026-09-10, 13 posts)

Developer Pawel Huryn (author of The Product Compass) runs Bug Hunt Bench: 105 real bugs deliberately injected into two production codebases, with frontier models (GPT-6, Claude, Grok, Gemini, DeepSeek, Kimi, GLM) asked to fix them in their agentic setups, blindly scored, one prompt per repo, results continuously updated on his GitHub leaderboard. This round's highlight is DeepSeek-V4.1-Flash's cost-effectiveness; Huryn is also retesting Kimi K3 and GLM at their highest reasoning tiers.

Confirmed

  • GPT-6 Astra supplementary scores: low tier 27/105, medium tier 34/105, published Sep 4-5 and viewable via leaderboard filters.
  • DeepSeek-V4.1-Flash (high tier): 24/105 fixed at max effort, total cost $1.80, 42.6 minutes.
  • Huryn's retest shows V4.1 Flash clearly beats V4 Pro: DeepSeek V4 Pro (max) fixed only 16/105 at $1.89 and 37 minutes.
  • Per secondhand accounts of the same bench: Claude Opus 5 (max) fixed 27 at $51.33; Grok 4.6 (max) and other flagships are also compared; Flash's $1.80 is 1/28 (3.5%) of Opus's cost.
  • Kimi K3 and GLM retests at highest reasoning tiers are in progress, results not yet posted.

Unconfirmed

  • The Opus/Grok comparison figures come from reposted summaries; the GitHub leaderboard is the authoritative source.

Why it matters

  • Real repos, injected bugs, blind scoring, and single-prompt setup make this closer to real-world debugging than synthetic benchmarks.
  • Flash fixing 24 bugs for $1.80, beating its own pricier V4 Pro, offers valuable selection data for budget-sensitive teams.
  • Continuous retesting may turn the leaderboard into a living reference for frontier coding models' bug-fixing ability.

Episode 11 · DeepSeek V4.1 Flash Confirmed to Center on YOCO Architecture (2026-09-10, 7 posts)

On Sept 10–11, commentators focused on the architecture of DeepSeek's V4.1 Flash. DeepSeek researcher Dong Li confirmed the model's soul is YOCO (You Only Cache Once), inheriting its order-of-magnitude advantages in prefill and KV cache; close readings of the paper also surfaced KV cache reuse across checkpoints, and the related techniques are fully open-sourced. TL;DR: V4.1 Flash's core gains come from YOCO's prefill/KV-cache optimizations, and DeepSeek is credited with pushing architecture-level innovation into the open-source community.

Confirmed

  • Dong Li (@donglixp) said V4.1 Flash's architecture centers on YOCO and that architectural innovation has entered its second half.
  • Astra Pro's paper deep-dive found KV cache reuse between checkpoints, with one commenter calling DeepSeek a "weird model factory."
  • The DeepSeek paper cites the 2024 YoCo paper as the main inspiration for its Causal-Encoder Decoder (CED) transformer module.
  • The new architecture's techniques are publicly open-sourced.

Unconfirmed

  • Chinese KOL MaxForAI's (unofficial) reading of the paper's diagrams suggests at least three structural changes, including a back-and-forth encoder-decoder layout and CSA2 cross-layer reuse; the real story is changes to the Transformer itself, not parameters or benchmarks.
  • Developer aHapBean argues the "Causal Encoder-Decoder" label overstates the novelty: V4.1 may still be a decoder-only LLM, better described via the YoCo framing.

Why it matters

  • @evijit notes DeepSeek keeps innovating at the architecture level, not just scale, and open-sourcing it raises the floor for the whole open-source ecosystem.
  • If order-of-magnitude KV cache optimization and cross-checkpoint reuse prove out, they could meaningfully cut training and deployment costs and directly improve inference economics.

Episode 12 · Leaked DeepSeek V4.1 Flash Tech Report Reveals KV Cache-Obsessed Flagship (2026-09-11, 44 posts)

A technical report for DeepSeek V4.1 Flash leaked online; @burkov cautions the release is not officially confirmed by DeepSeek. The report describes a flagship model built around KV cache compression: @stochasticchasm notes startling benchmark numbers, a 4x smaller KV cache than DSV4-Flash, and more stable training; @nrehiew published a long deep-dive series, with complementary analyses from Mhr1036 (via @demianai) and arjunkocher (via @teortaxesTex). Community consensus: systematic innovation across architecture, algorithms, training, and inference simultaneously cuts compute and memory costs for long context.

Confirmed

  • 552B-parameter MoE backbone with a 1M-token context window; Mhr1036 notes the "paying to keep context" second cost of long context, which this design compresses on both compute and memory; @stochasticchasm reports a 4x smaller KV cache vs DSV4-Flash
  • Core architecture is an asymmetric causal encoder-decoder (CED): 40 layers split into 20 encoder + 20 decoder; @nrehiew explains lower layers generate KV cache shared upward via per-layer projections — a YOCO-style decoder-decoder halving KV cache; arjunkocher confirms the asymmetric design and CSA2 attention; @stochasticchasm corrects: the decoder only periodically reads final encoder states, not standard cross-attention
  • Per Chris Alexiuk (via @altryne), prefill/decode separation is "baked" into the model weights, not just the serving stack
  • Global attention is upgraded compressed attention: three KV-reuse variants all run a sparse indexer over a small KV, then concatenate selected tokens with windowed KV; the indexer itself is hierarchical (full-mode first-stage candidate selection)
  • Head-wise Muon optimizer (including vision), a 196B-parameter Engram component with Sinkhorn balancing; sparse attention trained from scratch without warmup; @stochasticchasm notes no input/output optimizer parameters, weight decay only on the LM head, and suspects Muon is becoming the training default
  • KV cache is quantization-aware trained (QAT), explaining strong fp4 KV cache performance
  • Multimodal: 45T image-text tokens, in-house SigLIP-style encoder, modality-level load balancing (reminiscent of Chameleon per @nrehiew)
  • RL infrastructure (@nrehiew): dispatch strategies eliminating tail stalls, router replay from old checkpoints, off-policy controls for short completions, 40+ teacher OPDs; efficiency rated "Sol/Opus-level" with "insane" benchmark numbers
  • Sandbox & data (@nrehiew): in-house orchestrator, 2500 concurrent sandboxes per node per VM within NUMA domains; agent trajectory data explicitly inspired by real partner usage patterns, synthesized by multiple agents, with checkpoint merging
  • Inference (@nrehiew): heavily fused kernels (15 for prefill, 11 for decode); SWA cache kept per user/assistant message (minute-level); first model @nrehiew has seen with a numeric reasoning-effort parameter directly affecting length penalty
  • Training infra: parallel all-gather of one modality's features while another computes, context parallelism for long multi-image sequences, dedicated sparse-indexer clusters
  • Design almost solely targets inference efficiency: friendly to PD disagg with good utilization at large prefill batches; some commenters find v4's CSA/HCA dual compression-frequency design under-explained vs pure CSA2
  • @stochasticchasm observes DeepSeek appears to break its "V<N> = new pretrain" naming convention; some suggest this generation is a simplified V4

Unconfirmed

  • @burkov stresses the report is not officially confirmed by DeepSeek
  • @stochasticchasm's speculations linking the leaked architecture to rumored qwen-3.8-flash-next (51B n-gram parameters) and mechanisms like globally shared KV + per-layer SWA are personal conjecture

Why it matters

The community takeaway is that obsession with KV cache compression produced an exceptionally efficient flagship. As 1M-token contexts and long-horizon agent workloads drive up KV cache read/write and memory costs, V4.1 Flash demonstrates a systemic cost-reduction path — from architecture (asymmetric CED, KV reuse, baked-in PD separation) and algorithms (hierarchical sparse indexer, compressed attention, KV QAT) to training (Muon, Sinkhorn, router replay) and deployment (15/11 fused kernels, message-level SWA cache, numeric reasoning effort) — a valuable blueprint for long-context multimodal frontier models.

24 more related posts →

Episode 13 · DeepSeek V4.1 Flash tops Vals open-source index at $0.30 per run (2026-09-11, 2 posts)

DeepSeek V4.1 Flash ranks first among open-weight models on the Vals Index, beating Kimi K3 while costing only $0.30 per run — roughly 1/20 the price of competitors like Kimi K3 and GLM 5.3.

Episode 14 · DeepSeek launches V4.1-Flash: 552B backbone, 1M-token context (2026-09-11, 5 posts)

DeepSeek released DeepSeek-V4.1-Flash around September 11, the smallest member of its new architecture family, updating both its web app and mobile app with a "fast, expert, visual" positioning. It is a multimodal MoE with a 552B backbone (763B total parameters per Ollama's listing), native visual understanding, up to 1M-token context, and support for vision, tools, and thinking.

Confirmed

  • DeepSeek officially confirmed V4.1-Flash is the smallest model in the new family, promising higher capability, faster inference, and higher throughput, paving the way for larger models.
  • The model uses a Causal Encoder-Decoder architecture with 4x smaller KV cache.
  • Fireworks launched the model targeting coding and agent use cases; it is also available on the FLock API platform and Ollama (cloud mode).
  • Per APPSO, the backbone doubled from 284B to 552B while running faster: measured at 420–507 tok/s, with one user reporting 217 tok/s.

Unconfirmed

  • Parameter figures differ across sources: reposts cite 552B, while Ollama lists 763B total / 552B backbone; the official figure needs verification.
  • Claims such as "website-building cost 1/40 of Claude" and "Agent scores surpassing Claude" come from reposts and media, with unclear test conditions.

Why it matters

  • As the first small-size member of DeepSeek's new architecture family, its capability and cost profile sets the tone for larger models to come.
  • A 552B backbone combined with 1M-token context and 4x KV-cache compression poses direct competitive pressure on closed-source peers in long-context and agentic scenarios.

Episode 15 · DeepSeek V4.1 Flash scores 40 in AA index despite rising hallucinations (2026-09-11, 2 posts)

Artificial Analysis gives DeepSeek V4.1 Flash a score of 40, beating DeepSeek V4 Pro though trailing GLM-5.3-Flash. Its AA-Omniscience rose 9 points to -5.3, but hallucination rates increased.

Episode 16 · Inside KV-Cache Sharing: How DeepSeek CED and GLM 5.2 Differ (2026-09-11, 4 posts)

Developers break down KV-cache sharing designs: YOCO lets all layers in the model's second half share one KV cache; DeepSeek's CED applies a similar YOCO-style global branch per CSA2 layer, caching only encoder KV; GLM 5.2 scores per token while sharing block indices for flexibility and performance.

Episode 17 · DeepSeek 4.1 Flash Shows Big Gains in Kernel Engineering and Cost (2026-09-11, 2 posts)

DeepSeek's 4.1 Flash is praised for major improvements in kernel engineering, coding and 3D scene generation, with cache-hit pricing cut roughly sevenfold, while users say GLM-5.3 Flash lags far behind.

Episode 18 · SGLang delivers Day 0 support for DeepSeek V4.1 Flash, hitting 873 tok/s (2026-09-12, 2 posts)

SGLang announced Day 0 support for DeepSeek V4.1 Flash and pushed inference to 873 tok/s (BS=1) on 4×GB300 within 24 hours of the model's release, publishing an adaptation guide.